Uploaded by ceyale8443

Learning Bayesian Networks(Neapolitan, Richard)

advertisement
Learning Bayesian Networks
Richard E. Neapolitan
Northeastern Illinois University
Chicago, Illinois
In memory of my dad, a difficult but loving father, who raised me well.
ii
Contents
Preface
ix
I
1
Basics
1 Introduction to Bayesian Networks
1.1 Basics of Probability Theory . . . . . . . . . . . . . . . . . . . .
1.1.1 Probability Functions and Spaces . . . . . . . . . . . . . .
1.1.2 Conditional Probability and Independence . . . . . . . . .
1.1.3 Bayes’ Theorem . . . . . . . . . . . . . . . . . . . . . . .
1.1.4 Random Variables and Joint Probability Distributions . .
1.2 Bayesian Inference . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2.1 Random Variables and Probabilities in Bayesian Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2.2 A Definition of Random Variables and Joint Probability
Distributions for Bayesian Inference . . . . . . . . . . . .
1.2.3 A Classical Example of Bayesian Inference . . . . . . . . .
1.3 Large Instances / Bayesian Networks . . . . . . . . . . . . . . . .
1.3.1 The Difficulties Inherent in Large Instances . . . . . . . .
1.3.2 The Markov Condition . . . . . . . . . . . . . . . . . . . .
1.3.3 Bayesian Networks . . . . . . . . . . . . . . . . . . . . . .
1.3.4 A Large Bayesian Network . . . . . . . . . . . . . . . . .
1.4 Creating Bayesian Networks Using Causal Edges . . . . . . . . .
1.4.1 Ascertaining Causal Influences Using Manipulation . . . .
1.4.2 Causation and the Markov Condition . . . . . . . . . . .
3
5
6
9
12
13
20
2 More DAG/Probability Relationships
2.1 Entailed Conditional Independencies . . . . . . . . . . . .
2.1.1 Examples of Entailed Conditional Independencies .
2.1.2 d-Separation . . . . . . . . . . . . . . . . . . . . .
2.1.3 Finding d-Separations . . . . . . . . . . . . . . . .
2.2 Markov Equivalence . . . . . . . . . . . . . . . . . . . . .
2.3 Entailing Dependencies with a DAG . . . . . . . . . . . .
2.3.1 Faithfulness . . . . . . . . . . . . . . . . . . . . . .
65
66
66
70
76
84
92
95
iii
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
20
24
27
29
29
31
40
43
43
44
51
iv
CONTENTS
2.4
2.5
2.6
II
2.3.2 Embedded Faithfulness . . . . . . . . . . . . . .
Minimality . . . . . . . . . . . . . . . . . . . . . . . . .
Markov Blankets and Boundaries . . . . . . . . . . . . .
More on Causal DAGs . . . . . . . . . . . . . . . . . . .
2.6.1 The Causal Minimality Assumption . . . . . . .
2.6.2 The Causal Faithfulness Assumption . . . . . . .
2.6.3 The Causal Embedded Faithfulness Assumption
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Inference
3 Inference: Discrete Variables
3.1 Examples of Inference . . . . . . . . . . . . . . . .
3.2 Pearl’s Message-Passing Algorithm . . . . . . . . .
3.2.1 Inference in Trees . . . . . . . . . . . . . . .
3.2.2 Inference in Singly-Connected Networks . .
3.2.3 Inference in Multiply-Connected Networks .
3.2.4 Complexity of the Algorithm . . . . . . . .
3.3 The Noisy OR-Gate Model . . . . . . . . . . . . .
3.3.1 The Model . . . . . . . . . . . . . . . . . .
3.3.2 Doing Inference With the Model . . . . . .
3.3.3 Further Models . . . . . . . . . . . . . . . .
3.4 Other Algorithms that Employ the DAG . . . . . .
3.5 The SPI Algorithm . . . . . . . . . . . . . . . . . .
3.5.1 The Optimal Factoring Problem . . . . . .
3.5.2 Application to Probabilistic Inference . . .
3.6 Complexity of Inference . . . . . . . . . . . . . . .
3.7 Relationship to Human Reasoning . . . . . . . . .
3.7.1 The Causal Network Model . . . . . . . . .
3.7.2 Studies Testing the Causal Network Model
99
104
108
110
110
111
112
121
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
123
124
126
127
142
153
155
156
156
160
161
161
162
163
168
170
171
171
173
4 More Inference Algorithms
4.1 Continuous Variable Inference . . . . . . . . . . . . . . . . . . .
4.1.1 The Normal Distribution . . . . . . . . . . . . . . . . .
4.1.2 An Example Concerning Continuous Variables . . . . .
4.1.3 An Algorithm for Continuous Variables . . . . . . . . .
4.2 Approximate Inference . . . . . . . . . . . . . . . . . . . . . . .
4.2.1 A Brief Review of Sampling . . . . . . . . . . . . . . . .
4.2.2 Logic Sampling . . . . . . . . . . . . . . . . . . . . . . .
4.2.3 Likelihood Weighting . . . . . . . . . . . . . . . . . . . .
4.3 Abductive Inference . . . . . . . . . . . . . . . . . . . . . . . .
4.3.1 Abductive Inference in Bayesian Networks . . . . . . . .
4.3.2 A Best-First Search Algorithm for Abductive Inference .
.
.
.
.
.
.
.
.
.
.
.
181
181
182
183
185
205
205
211
217
221
221
224
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
CONTENTS
5 Influence Diagrams
5.1 Decision Trees . . . . . . . . . . . . . . . . . . .
5.1.1 Simple Examples . . . . . . . . . . . . .
5.1.2 Probabilities, Time, and Risk Attitudes
5.1.3 Solving Decision Trees . . . . . . . . . .
5.1.4 More Examples . . . . . . . . . . . . . .
5.2 Influence Diagrams . . . . . . . . . . . . . . . .
5.2.1 Representing with Influence Diagrams .
5.2.2 Solving Influence Diagrams . . . . . . .
5.3 Dynamic Networks . . . . . . . . . . . . . . . .
5.3.1 Dynamic Bayesian Networks . . . . . .
5.3.2 Dynamic Influence Diagrams . . . . . .
III
v
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Learning
239
239
239
242
245
245
259
259
266
272
272
279
291
6 Parameter Learning: Binary Variables
6.1 Learning a Single Parameter . . . . . . . . . . . . . . . . . . . . .
6.1.1 Probability Distributions of Relative Frequencies . . . . .
6.1.2 Learning a Relative Frequency . . . . . . . . . . . . . . .
6.2 More on the Beta Density Function . . . . . . . . . . . . . . . . .
6.2.1 Non-integral Values of a and b . . . . . . . . . . . . . . .
6.2.2 Assessing the Values of a and b . . . . . . . . . . . . . . .
6.2.3 Why the Beta Density Function? . . . . . . . . . . . . . .
6.3 Computing a Probability Interval . . . . . . . . . . . . . . . . . .
6.4 Learning Parameters in a Bayesian Network . . . . . . . . . . . .
6.4.1 Urn Examples . . . . . . . . . . . . . . . . . . . . . . . .
6.4.2 Augmented Bayesian Networks . . . . . . . . . . . . . . .
6.4.3 Learning Using an Augmented Bayesian Network . . . . .
6.4.4 A Problem with Updating; Using an Equivalent Sample
Size . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.5 Learning with Missing Data Items . . . . . . . . . . . . . . . . .
6.5.1 Data Items Missing at Random . . . . . . . . . . . . . . .
6.5.2 Data Items Missing Not at Random . . . . . . . . . . . .
6.6 Variances in Computed Relative Frequencies . . . . . . . . . . . .
6.6.1 A Simple Variance Determination . . . . . . . . . . . . .
6.6.2 The Variance and Equivalent Sample Size . . . . . . . . .
6.6.3 Computing Variances in Larger Networks . . . . . . . . .
6.6.4 When Do Variances Become Large? . . . . . . . . . . . .
293
294
294
303
310
311
313
315
319
323
323
331
336
7 More Parameter Learning
7.1 Multinomial Variables . . . . . . . . . . . . . . . . .
7.1.1 Learning a Single Parameter . . . . . . . . .
7.1.2 More on the Dirichlet Density Function . . .
7.1.3 Computing Probability Intervals and Regions
7.1.4 Learning Parameters in a Bayesian Network .
381
381
381
388
389
392
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
348
357
358
363
364
364
366
372
373
vi
CONTENTS
7.1.5 Learning with Missing Data Items . . . . . .
7.1.6 Variances in Computed Relative Frequencies
7.2 Continuous Variables . . . . . . . . . . . . . . . . . .
7.2.1 Normally Distributed Variable . . . . . . . .
7.2.2 Multivariate Normally Distributed Variables
7.2.3 Gaussian Bayesian Networks . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
398
398
398
399
413
425
8 Bayesian Structure Learning
8.1 Learning Structure: Discrete Variables . . . . . . . . . . . . . . .
8.1.1 Schema for Learning Structure . . . . . . . . . . . . . . .
8.1.2 Procedure for Learning Structure . . . . . . . . . . . . . .
8.1.3 Learning From a Mixture of Observational and Experimental Data. . . . . . . . . . . . . . . . . . . . . . . . . .
8.1.4 Complexity of Structure Learning . . . . . . . . . . . . .
8.2 Model Averaging . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.3 Learning Structure with Missing Data . . . . . . . . . . . . . . .
8.3.1 Monte Carlo Methods . . . . . . . . . . . . . . . . . . . .
8.3.2 Large-Sample Approximations . . . . . . . . . . . . . . .
8.4 Probabilistic Model Selection . . . . . . . . . . . . . . . . . . . .
8.4.1 Probabilistic Models . . . . . . . . . . . . . . . . . . . . .
8.4.2 The Model Selection Problem . . . . . . . . . . . . . . . .
8.4.3 Using the Bayesian Scoring Criterion for Model Selection
8.5 Hidden Variable DAG Models . . . . . . . . . . . . . . . . . . . .
8.5.1 Models Containing More Conditional Independencies than
DAG Models . . . . . . . . . . . . . . . . . . . . . . . . .
8.5.2 Models Containing the Same Conditional Independencies
as DAG Models . . . . . . . . . . . . . . . . . . . . . . . .
8.5.3 Dimension of Hidden Variable DAG Models . . . . . . . .
8.5.4 Number of Models and Hidden Variables . . . . . . . . . .
8.5.5 Efficient Model Scoring . . . . . . . . . . . . . . . . . . .
8.6 Learning Structure: Continuous Variables . . . . . . . . . . . . .
8.6.1 The Density Function of D . . . . . . . . . . . . . . . . .
8.6.2 The Density function of D Given a DAG pattern . . . . .
8.7 Learning Dynamic Bayesian Networks . . . . . . . . . . . . . . .
441
441
442
445
9 Approximate Bayesian Structure Learning
9.1 Approximate Model Selection . . . . . . . . . . . . . . .
9.1.1 Algorithms that Search over DAGs . . . . . . . .
9.1.2 Algorithms that Search over DAG Patterns . . .
9.1.3 An Algorithm Assuming Missing Data or Hidden
9.2 Approximate Model Averaging . . . . . . . . . . . . . .
9.2.1 A Model Averaging Example . . . . . . . . . . .
9.2.2 Approximate Model Averaging Using MCMC . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
449
450
451
452
453
462
468
468
472
473
476
477
479
484
486
487
491
491
495
505
511
. . . . . 511
. . . . . 513
. . . . . 518
Variables 529
. . . . . 531
. . . . . 532
. . . . . 533
CONTENTS
vii
10 Constraint-Based Learning
541
10.1 Algorithms Assuming Faithfulness . . . . . . . . . . . . . . . . . 542
10.1.1 Simple Examples . . . . . . . . . . . . . . . . . . . . . . . 542
10.1.2 Algorithms for Determining DAG patterns . . . . . . . . 545
10.1.3 Determining if a Set Admits a Faithful DAG Representation552
10.1.4 Application to Probability . . . . . . . . . . . . . . . . . . 560
10.2 Assuming Only Embedded Faithfulness . . . . . . . . . . . . . . 561
10.2.1 Inducing Chains . . . . . . . . . . . . . . . . . . . . . . . 562
10.2.2 A Basic Algorithm . . . . . . . . . . . . . . . . . . . . . . 568
10.2.3 Application to Probability . . . . . . . . . . . . . . . . . . 590
10.2.4 Application to Learning Causal Influences1 . . . . . . . . 591
10.3 Obtaining the d-separations . . . . . . . . . . . . . . . . . . . . . 599
10.3.1 Discrete Bayesian Networks . . . . . . . . . . . . . . . . . 600
10.3.2 Gaussian Bayesian Networks . . . . . . . . . . . . . . . . 603
10.4 Relationship to Human Reasoning . . . . . . . . . . . . . . . . . 604
10.4.1 Background Theory . . . . . . . . . . . . . . . . . . . . . 604
10.4.2 A Statistical Notion of Causality . . . . . . . . . . . . . . 606
11 More Structure Learning
11.1 Comparing the Methods . . . . . . . . . . . . .
11.1.1 A Simple Example . . . . . . . . . . . .
11.1.2 Learning College Attendance Influences
11.1.3 Conclusions . . . . . . . . . . . . . . . .
11.2 Data Compression Scoring Criteria . . . . . . .
11.3 Parallel Learning of Bayesian Networks . . . .
11.4 Examples . . . . . . . . . . . . . . . . . . . . .
11.4.1 Structure Learning . . . . . . . . . . . .
11.4.2 Inferring Causal Relationships . . . . .
IV
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Applications
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
617
617
618
620
623
624
624
624
625
633
647
12 Applications
649
12.1 Applications Based on Bayesian Networks . . . . . . . . . . . . . 649
12.2 Beyond Bayesian networks . . . . . . . . . . . . . . . . . . . . . . 655
Bibliography
657
Index
686
1 The
relationships in the examples in this section are largely fictitious.
viii
CONTENTS
Preface
Bayesian networks are graphical structures for representing the probabilistic
relationships among a large number of variables and doing probabilistic inference
with those variables. During the 1980’s, a good deal of related research was done
on developing Bayesian networks (belief networks, causal networks, influence
diagrams), algorithms for performing inference with them, and applications that
used them. However, the work was scattered throughout research articles. My
purpose in writing the 1990 text Probabilistic Reasoning in Expert Systems was
to unify this research and establish a textbook and reference for the field which
has come to be known as ‘Bayesian networks.’ The 1990’s saw the emergence
of excellent algorithms for learning Bayesian networks from data. However,
by 2000 there still seemed to be no accessible source for ‘learning Bayesian
networks.’ Similar to my purpose a decade ago, the goal of this text is to
provide such a source.
In order to make this text a complete introduction to Bayesian networks,
I discuss methods for doing inference in Bayesian networks and influence diagrams. However, there is no effort to be exhaustive in this discussion. For
example, I give the details of only two algorithms for exact inference with discrete variables, namely Pearl’s message passing algorithm and D’Ambrosio and
Li’s symbolic probabilistic inference algorithm. It may seem odd that I present
Pearl’s algorithm, since it is one of the oldest. I have two reasons for doing
this: 1) Pearl’s algorithm corresponds to a model of human causal reasoning,
which is discussed in this text; and 2) Pearl’s algorithm extends readily to an
algorithm for doing inference with continuous variables, which is also discussed
in this text.
The content of the text is as follows. Chapters 1 and 2 cover basics. Specifically, Chapter 1 provides an introduction to Bayesian networks; and Chapter 2
discusses further relationships between DAGs and probability distributions such
as d-separation, the faithfulness condition, and the minimality condition. Chapters 3-5 concern inference. Chapter 3 covers Pearl’s message-passing algorithm,
D’Ambrosio and Li’s symbolic probabilistic inference, and the relationship of
Pearl’s algorithm to human causal reasoning. Chapter 4 shows an algorithm for
doing inference with continuous variable, an approximate inference algorithm,
and finally an algorithm for abductive inference (finding the most probable
explanation). Chapter 5 discusses influence diagrams, which are Bayesian networks augmented with decision nodes and a value node, and dynamic Bayesian
ix
x
PREFACE
networks and influence diagrams. Chapters 6-10 address learning. Chapters
6 and 7 concern parameter learning. Since the notation for these learning algorithm is somewhat arduous, I introduce the algorithms by discussing binary
variables in Chapter 6. I then generalize to multinomial variables in Chapter 7.
Furthermore, in Chapter 7 I discuss learning parameters when the variables are
continuous. Chapters 8, 9, and 10 concern structure learning. Chapter 8 shows
the Bayesian method for learning structure in the cases of both discrete and
continuous variables, while Chapter 9 discusses the constraint-based method
for learning structure. Chapter 10 compares the Bayesian and constraint-based
methods, and it presents several real-world examples of learning Bayesian networks. The text ends by referencing applications of Bayesian networks in Chapter 11.
This is a text on learning Bayesian networks; it is not a text on artificial
intelligence, expert systems, or decision analysis. However, since these are fields
in which Bayesian networks find application, they emerge frequently throughout
the text. Indeed, I have used the manuscript for this text in my course on expert
systems at Northeastern Illinois University. In one semester, I have found that
I can cover the core of the following chapters: 1, 2, 3, 5, 6, 7, 8, and 9.
I would like to thank those researchers who have provided valuable corrections, comments, and dialog concerning the material in this text. They include Bruce D’Ambrosio, David Maxwell Chickering, Gregory Cooper, Tom
Dean, Carl Entemann, John Erickson, Finn Jensen, Clark Glymour, Piotr
Gmytrasiewicz, David Heckerman, Xia Jiang, James Kenevan, Henry Kyburg,
Kathryn Blackmond Laskey, Don Labudde, David Madigan, Christopher Meek,
Paul-André Monney, Scott Morris, Peter Norvig, Judea Pearl, Richard Scheines,
Marco Valtorta, Alex Wolpert, and Sandy Zabell. I thank Sue Coyle for helping
me draw the cartoon containing the robots.
Part I
Basics
1
Chapter 1
Introduction to Bayesian
Networks
Consider the situation where one feature of an entity has a direct influence on
another feature of that entity. For example, the presence or absence of a disease
in a human being has a direct influence on whether a test for that disease turns
out positive or negative. For decades, Bayes’ theorem has been used to perform
probabilistic inference in this situation. In the current example, we would use
that theorem to compute the conditional probability of an individual having a
disease when a test for the disease came back positive. Consider next the situation where several features are related through inference chains. For example,
whether or not an individual has a history of smoking has a direct influence
both on whether or not that individual has bronchitis and on whether or not
that individual has lung cancer. In turn, the presence or absence of each of these
diseases has a direct influence on whether or not the individual experiences fatigue. Also, the presence or absence of lung cancer has a direct influence on
whether or not a chest X-ray is positive. In this situation, we would want to do
probabilistic inference involving features that are not related via a direct influence. We would want to determine, for example, the conditional probabilities
both of bronchitis and of lung cancer when it is known an individual smokes, is
fatigued, and has a positive chest X-ray. Yet bronchitis has no direct influence
(indeed no influence at all) on whether a chest X-ray is positive. Therefore,
these conditional probabilities cannot be computed using a simple application
of Bayes’ theorem. There is a straightforward algorithm for computing them,
but the probability values it requires are not ordinarily accessible; furthermore,
the algorithm has exponential space and time complexity.
Bayesian networks were developed to address these difficulties. By exploiting
conditional independencies entailed by influence chains, we are able to represent
a large instance in a Bayesian network using little space, and we are often able
to perform probabilistic inference among the features in an acceptable amount
of time. In addition, the graphical nature of Bayesian networks gives us a much
3
4
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
P(h1) = .2
H
P(b1|h1) = .25
P(b1|h2) = .05
B
L
P(l1|h1) = .003
P(l1|h2) = .00005
F
P(f1|b1,l1) = .75
P(f1|b1,l2) = .10
P(f1|b2,l1) = .5
P(f1|b2,l2) = .05
C
P(c1|l1) = .6
P(c1|l2) = .02
Figure 1.1: A Bayesian nework.
better intuitive grasp of the relationships among the features.
Figure 1.1 shows a Bayesian network representing the probabilistic relationships among the features just discussed. The values of the features in that
network represent the following:
Feature
H
B
L
F
C
Value
h1
h2
b1
b2
l1
l2
f1
f2
c1
c2
When the Feature Takes this Value
There is a history of smoking
There is no history of smoking
Bronchitis is present
Bronchitis is absent
Lung cancer is present
Lung cancer is absent
Fatigue is present
Fatigue is absent
Chest X-ray is positive
Chest X-ray is negative
This Bayesian network is discussed in Example 1.32 in Section 1.3.3 after we
provide the theory of Bayesian networks. Presently, we only use it to illustrate
the nature and use of Bayesian networks. First, in this Bayesian network (called
a causal network) the edges represent direct influences. For example, there is
an edge from H to L because a history of smoking has a direct influence on the
presence of lung cancer, and there is an edge from L to C because the presence
of lung cancer has a direct influence on the result of a chest X-ray. There is no
1.1. BASICS OF PROBABILITY THEORY
5
edge from H to C because a history of smoking has an influence on the result
of a chest X-ray only through its influence on the presence of lung cancer. One
way to construct Bayesian networks is by creating edges that represent direct
influences as done here; however, there are other ways. Second, the probabilities
in the network are the conditional probabilities of the values of each feature given
every combination of values of the feature’s parents in the network, except in the
case of roots they are prior probabilities. Third, probabilistic inference among
the features can be accomplished using the Bayesian network. For example, we
can compute the conditional probabilities both of bronchitis and of lung cancer
when it is known an individual smokes, is fatigued, and has a positive chest
X-ray. This Bayesian network is discussed again in Chapter 3 when we develop
algorithms that do this inference.
The focus of this text is on learning Bayesian networks from data. For
example, given we had values of the five features just discussed (smoking history, bronchitis, lung cancer, fatigue, and chest X-ray) for a large number of
individuals, the learning algorithms we develop might construct the Bayesian
network in Figure 1.1. However, to make it a complete introduction to Bayesian
networks, it does include a brief overview of methods for doing inference in
Bayesian networks and using Bayesian networks to make decisions. Chapters 1
and 2 cover properties of Bayesian networks which we need in order to discuss
both inference and learning. Chapters 3-5 concern methods for doing inference
in Bayesian networks. Methods for learning Bayesian networks from data are
discussed in Chapters 6-11. A number of successful experts systems (systems
which make the judgements of an expert) have been developed which are based
on Bayesian networks. Furthermore, Bayesian networks have been used to learn
causal influences from data. Chapter 12 references some of these real-world applications. To see the usefulness of Bayesian networks, you may wish to review
that chapter before proceeding.
This chapter introduces Bayesian networks. Section 1.1 reviews basic concepts in probability. Next, Section 1.2 discusses Bayesian inference and illustrates the classical way of using Bayes’ theorem when there are only two features. Section 1.3 shows the problem in representing large instances and introduces Bayesian networks as a solution to this problem. Finally, we discuss how
Bayesian networks can often be constructed using causal edges.
1.1
Basics of Probability Theory
The concept of probability has a rich and diversified history that includes many
different philosophical approaches. Notable among these approaches include the
notions of probability as a ratio, as a relative frequency, and as a degree of belief.
Next we review the probability calculus and, via examples, illustrate these three
approaches and how they are related.
6
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
1.1.1
Probability Functions and Spaces
In 1933 A.N. Kolmogorov developed the set-theoretic definition of probability,
which serves as a mathematical foundation for all applications of probability.
We start by providing that definition.
Probability theory has to do with experiments that have a set of distinct
outcomes. Examples of such experiments include drawing the top card from a
deck of 52 cards with the 52 outcomes being the 52 different faces of the cards;
flipping a two-sided coin with the two outcomes being ‘heads’ and ‘tails’; picking
a person from a population and determining whether the person is a smoker
with the two outcomes being ‘smoker’ and ‘non-smoker’; picking a person from
a population and determining whether the person has lung cancer with the
two outcomes being ‘having lung cancer’ and ‘not having lung cancer’; after
identifying 5 levels of serum calcium, picking a person from a population and
determining the individual’s serum calcium level with the 5 outcomes being
each of the 5 levels; picking a person from a population and determining the
individual’s serum calcium level with the infinite number of outcomes being
the continuum of possible calcium levels. The last two experiments illustrate
two points. First, the experiment is not well-defined until we identify a set of
outcomes. The same act (picking a person and measuring that person’s serum
calcium level) can be associated with many different experiments, depending on
what we consider a distinct outcome. Second, the set of outcomes can be infinite.
Once an experiment is well-defined, the collection of all outcomes is called the
sample space. Mathematically, a sample space is a set and the outcomes are
the elements of the set. To keep this review simple, we restrict ourselves to finite
sample spaces in what follows (You should consult a mathematical probability
text such as [Ash, 1970] for a discussion of infinite sample spaces.). In the case
of a finite sample space, every subset of the sample space is called an event. A
subset containing exactly one element is called an elementary event. Once a
sample space is identified, a probability function is defined as follows:
Definition 1.1 Suppose we have a sample space Ω containing n distinct elements. That is,
Ω = {e1 , e2 , . . . en }.
A function that assigns a real number P (E) to each event E ⊆ Ω is called
a probability function on the set of subsets of Ω if it satisfies the following
conditions:
1. 0 ≤ P ({ei }) ≤ 1
for 1 ≤ i ≤ n.
2. P ({e1 }) + P ({e2 }) + . . . + P ({en }) = 1.
3. For each event E = {ei1 , ei2 , . . . eik } that is not an elementary event,
P (E) = P ({ei1 }) + P ({ei2 }) + . . . + P ({eik }).
The pair (Ω, P ) is called a probability space.
1.1. BASICS OF PROBABILITY THEORY
7
We often just say P is a probability function on Ω rather than saying on the
set of subsets of Ω.
Intuition for probability functions comes from considering games of chance
as the following example illustrates.
Example 1.1 Let the experiment be drawing the top card from a deck of 52
cards. Then Ω contains the faces of the 52 cards, and using the principle of
indifference, we assign P ({e}) = 1/52 for each e ∈ Ω. Therefore, if we let kh
and ks stand for the king of hearts and king of spades respectively, P ({kh}) =
1/52, P ({ks}) = 1/52, and P ({kh, ks}) = P ({kh}) + P ({ks}) = 1/26.
The principle of indifference (a term popularized by J.M. Keynes in 1921)
says elementary events are to be considered equiprobable if we have no reason
to expect or prefer one over the other. According to this principle, when there
are n elementary events the probability of each of them is the ratio 1/n. This
is the way we often assign probabilities in games of chance, and a probability
so assigned is called a ratio.
The following example shows a probability that cannot be computed using
the principle of indifference.
Example 1.2 Suppose we toss a thumbtack and consider as outcomes the two
ways it could land. It could land on its head, which we will call ‘heads’, or
it could land with the edge of the head and the end of the point touching the
ground, which we will call ‘tails’. Due to the lack of symmetry in a thumbtack,
we would not assign a probability of 1/2 to each of these events. So how can
we compute the probability? This experiment can be repeated many times. In
1919 Richard von Mises developed the relative frequency approach to probability
which says that, if an experiment can be repeated many times, the probability of
any one of the outcomes is the limit, as the number of trials approach infinity,
of the ratio of the number of occurrences of that outcome to the total number of
trials. For example, if m is the number of trials,
P ({heads}) = lim
m→∞
#heads
.
m
So, if we tossed the thumbtack 10, 000 times and it landed heads 3373 times, we
would estimate the probability of heads to be about .3373.
Probabilities obtained using the approach in the previous example are called
relative frequencies. According to this approach, the probability obtained is
not a property of any one of the trials, but rather it is a property of the entire
sequence of trials. How are these probabilities related to ratios? Intuitively,
we would expect if, for example, we repeatedly shuffled a deck of cards and
drew the top card, the ace of spades would come up about one out of every 52
times. In 1946 J. E. Kerrich conducted many such experiments using games of
chance in which the principle of indifference seemed to apply (e.g. drawing a
card from a deck). His results indicated that the relative frequency does appear
to approach a limit and that limit is the ratio.
8
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
The next example illustrates a probability that cannot be obtained either
with ratios or with relative frequencies.
Example 1.3 If you were going to bet on an upcoming basketball game between
the Chicago Bulls and the Detroit Pistons, you would want to ascertain how
probable it was that the Bulls would win. This probability is certainly not a
ratio, and it is not a relative frequency because the game cannot be repeated
many times under the exact same conditions (Actually, with your knowledge
about the conditions the same.). Rather the probability only represents your
belief concerning the Bulls chances of winning. Such a probability is called a
degree of belief or subjective probability. There are a number of ways
for ascertaining such probabilities. One of the most popular methods is the
following, which was suggested by D. V. Lindley in 1985. This method says an
individual should liken the uncertain outcome to a game of chance by considering
an urn containing white and black balls. The individual should determine for
what fraction of white balls the individual would be indifferent between receiving
a small prize if the uncertain outcome happened (or turned out to be true) and
receiving the same small prize if a white ball was drawn from the urn. That
fraction is the individual’s probability of the outcome. Such a probability can be
constructed using binary cuts. If, for example, you were indifferent when the
fraction was .75, for you P ({bullswin}) = .75. If I were indifferent when the
fraction was .6, for me P ({bullswin}) = .6. Neither of us is right or wrong.
Subjective probabilities are unlike ratios and relative frequencies in that they do
not have objective values upon which we all must agree. Indeed, that is why they
are called subjective.
Neapolitan [1996] discusses the construction of subjective probabilities further. In this text, by probability we ordinarily mean a degree of belief. When
we are able to compute ratios or relative frequencies, the probabilities obtained
agree with most individuals’ beliefs. For example, most individuals would assign
a subjective probability of 1/13 to the top card being an ace because they would
be indifferent between receiving a small prize if it were the ace and receiving
that same small prize if a white ball were drawn from an urn containing one
white ball out of 13 total balls.
The following example shows a subjective probability more relevant to applications of Bayesian networks.
Example 1.4 After examining a patient and seeing the result of the patient’s
chest X-ray, Dr. Gloviak decides the probability that the patient has lung cancer
is .9. This probability is Dr. Gloviak’s subjective probability of that outcome.
Although a physician may use estimates of relative frequencies (such as the
fraction of times individuals with lung cancer have positive chest X-rays) and
experience diagnosing many similar patients to arrive at the probability, it is
still assessed subjectively. If asked, Dr. Gloviak may state that her subjective
probability is her estimate of the relative frequency with which patients, who
have these exact same symptoms, have lung cancer. However, there is no reason
to believe her subjective judgement will converge, as she continues to diagnose
1.1. BASICS OF PROBABILITY THEORY
9
patients with these exact same symptoms, to the actual relative frequency with
which they have lung cancer.
It is straightforward to prove the following theorem concerning probability
spaces.
Theorem 1.1 Let (Ω, P ) be a probability space. Then
1. P (Ω) = 1.
2. 0 ≤ P (E) ≤ 1
for every E ⊆ Ω.
3. For E and F ⊆ Ω such that E ∩ F = ∅,
P (E ∪ F) = P (E) + P (F).
Proof. The proof is left as an exercise.
The conditions in this theorem were labeled the axioms of probability
theory by A.N. Kolmogorov in 1933. When Condition (3) is replaced by infinitely countable additivity, these conditions are used to define a probability
space in mathematical probability texts.
Example 1.5 Suppose we draw the top card from a deck of cards. Denote by
Queen the set containing the 4 queens and by King the set containing the 4 kings.
Then
P (Queen ∪ King) = P (Queen) + P (King) = 1/13 + 1/13 = 2/13
because Queen ∩ King = ∅. Next denote by Spade the set containing the 13
spades. The sets Queen and Spade are not disjoint; so their probabilities are not
additive. However, it is not hard to prove that, in general,
P (E ∪ F) = P (E) + P (F) − P (E ∩ F).
So
P (Queen ∪ Spade) = P (Queen) + P (Spade) − P (Queen ∩ Spade)
1
1
4
1
+ −
= .
=
13 4 52
13
1.1.2
Conditional Probability and Independence
We have yet to discuss one of the most important concepts in probability theory,
namely conditional probability. We do that next.
Definition 1.2 Let E and F be events such that P (F) 6= 0. Then the conditional probability of E given F, denoted P (E|F), is given by
P (E|F) =
P (E ∩ F)
.
P (F)
10
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
The initial intuition for conditional probability comes from considering probabilities that are ratios. In the case of ratios, P (E|F), as defined above, is the
fraction of items in F that are also in E. We show this as follows. Let n be the
number of items in the sample space, nF be the number of items in F, and nEF
be the number of items in E ∩ F. Then
nEF /n
nEF
P (E ∩ F)
=
=
,
P (F)
nF /n
nF
which is the fraction of items in F that are also in E. As far as meaning, P (E|F)
means the probability of E occurring given that we know F has occurred.
Example 1.6 Again consider drawing the top card from a deck of cards, let
Queen be the set of the 4 queens, RoyalCard be the set of the 12 royal cards, and
Spade be the set of the 13 spades. Then
P (Queen) =
P (Queen|RoyalCard) =
P (Queen|Spade) =
1
13
1/13
1
P (Queen ∩ RoyalCard)
=
=
P (RoyalCard)
3/13
3
1/52
1
P (Queen ∩ Spade)
=
= .
P (Spade)
1/4
13
Notice in the previous example that P (Queen|Spade) = P (Queen). This
means that finding out the card is a spade does not make it more or less probable
that it is a queen. That is, the knowledge of whether it is a spade is irrelevant
to whether it is a queen. We say that the two events are independent in this
case, which is formalized in the following definition.
Definition 1.3 Two events E and F are independent if one of the following
hold:
1. P (E|F) = P (E)
and
P (E) 6= 0, P (F) 6= 0.
2. P (E) = 0 or P (F) = 0.
Notice that the definition states that the two events are independent even
though it is based on the conditional probability of E given F. The reason is
that independence is symmetric. That is, if P (E) 6= 0 and P (F) 6= 0, then
P (E|F) = P (E) if and only if P (F|E) = P (F). It is straightforward to prove that
E and F are independent if and only if P (E ∩ F) = P (E)P (F).
The following example illustrates an extension of the notion of independence.
Example 1.7 Let E = {kh, ks, qh}, F = {kh, kc, qh}, G = {kh, ks, kc, kd},
where kh means the king of hearts, ks means the king of spades, etc. Then
P (E) =
P (E|F) =
3
52
2
3
1.1. BASICS OF PROBABILITY THEORY
P (E|G) =
P (E|F ∩ G) =
11
2
1
=
4
2
1
.
2
So E and F are not independent, but they are independent once we condition on
G.
In the previous example, E and F are said to be conditionally independent
given G. Conditional independence is very important in Bayesian networks and
will be discussed much more in the sections that follow. Presently, we have the
definition that follows and another example.
Definition 1.4 Two events E and F are conditionally independent given G
if P (G) 6= 0 and one of the following holds:
1. P (E|F ∩ G) = P (E|G)
and
P (E|G) 6= 0, P (F|G) 6= 0.
2. P (E|G) = 0 or P (F|G) = 0.
Another example of conditional independence follows.
Example 1.8 Let Ω be the set of all objects in Figure 1.2. Suppose we assign
a probability of 1/13 to each object, and let Black be the set of all black objects,
White be the set of all white objects, Square be the set of all square objects, and
One be the set of all objects containing a ‘1’. We then have
P (One) =
P (One|Square) =
P (One|Black) =
P (One|Square ∩ Black) =
P (One|White) =
P (One|Square ∩ White) =
5
13
3
8
1
3
=
9
3
1
2
=
6
3
1
2
=
4
2
1
.
2
So One and Square are not independent, but they are conditionally independent
given Black and given White.
Next we discuss a very useful rule involving conditional probabilities. Suppose we have n events E1 , E2 , . . . En such that Ei ∩ Ej = ∅ for i 6= j and
12
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
1
1
2
2
1
2
1
2
2
2
1
2
2
Figure 1.2: Containing a ‘1’ and being a square are not independent, but they
are conditionally independent given the object is black and given it is white.
E1 ∪ E2 ∪ . . . ∪ En = Ω. Such events are called mutually exclusive and
exhaustive. Then the law of total probability says for any other event F,
P (F) =
n
X
P (F ∩ Ei ).
(1.1)
i=1
If P (Ei ) 6= 0, then P (F ∩ Ei ) = P (F|Ei )P (Ei ). Therefore, if P (Ei ) 6= 0 for all i,
the law is often applied in the following form:
P (F) =
n
X
P (F|Ei )P (Ei ).
(1.2)
i=1
It is straightforward to derive both the axioms of probability theory and
the rule for conditional probability when probabilities are ratios. However,
they can also be derived in the relative frequency and subjectivistic frameworks
(See [Neapolitan, 1990].). These derivations make the use of probability theory
compelling for handling uncertainty.
1.1.3
Bayes’ Theorem
For decades conditional probabilities of events of interest have been computed
from known probabilities using Bayes’ theorem. We develop that theorem next.
Theorem 1.2 (Bayes) Given two events E and F such that P (E) 6= 0 and
P (F) 6= 0, we have
P (F|E)P (E)
.
(1.3)
P (E|F) =
P (F)
Furthermore, given n mutually exclusive and exhaustive events E1 , E2 , . . . En
such that P (Ei ) 6= 0 for all i, we have for 1 ≤ i ≤ n,
P (Ei |F) =
P (F|Ei )P (Ei )
.
P (F|E1 )P (E1 ) + P (F|E2 )P (E2 ) + · · · P (F|En )P (En )
(1.4)
1.1. BASICS OF PROBABILITY THEORY
13
Proof. To obtain Equality 1.3, we first use the definition of conditional probability as follows:
P (E|F) =
P (E ∩ F)
P (F)
and
P (F|E) =
P (F ∩ E)
.
P (E)
Next we multiply each of these equalities by the denominator on its right side to
show that
P (E|F)P (F) = P (F|E)P (E)
because they both equal P (E ∩ F). Finally, we divide this last equality by P (F)
to obtain our result.
To obtain Equality 1.4, we place the expression for F, obtained using the rule
of total probability (Equality 1.2), in the denominator of Equality 1.3.
Both of the formulas in the preceding theorem are called Bayes’ theorem
because they were originally developed by Thomas Bayes (published in 1763).
The first enables us to compute P (E|F) if we know P (F|E), P (E), and P (F), while
the second enables us to compute P (Ei |F) if we know P (F|Ej ) and P (Ej ) for
1 ≤ j ≤ n. Computing a conditional probability using either of these formulas
is called Bayesian inference. An example of Bayesian inference follows:
Example 1.9 Let Ω be the set of all objects in Figure 1.2, and assign each
object a probability of 1/13. Let One be the set of all objects containing a 1, Two
be the set of all objects containing a 2, and Black be the set of all black objects.
Then according to Bayes’ Theorem,
P (One|Black) =
=
P (Black|One)P (One)
P (Black|One)P (One) + P (Black|Two)P (Two)
5
)
( 35 )( 13
1
= ,
3
5
6
8
3
( 5 )( 13 ) + ( 8 )( 13 )
which is the same value we get by computing P (One|Black) directly.
The previous example is not a very exciting application of Bayes’ Theorem
as we can just as easily compute P (One|Black) directly. Section 1.2 discusses
useful applications of Bayes’ Theorem.
1.1.4
Random Variables and Joint Probability Distributions
We have one final concept to discuss in this overview, namely that of a random
variable. The definition shown here is based on the set-theoretic definition of
probability given in Section 1.1.1. In Section 1.2.2 we provide an alternative
definition which is more pertinent to the way random variables are used in
practice.
Definition 1.5 Given a probability space (Ω, P ), a random variable X is a
function on Ω.
14
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
That is, a random variable assigns a unique value to each element (outcome)
in the sample space. The set of values random variable X can assume is called
the space of X. A random variable is said to be discrete if its space is finite
or countable. In general, we develop our theory assuming the random variables
are discrete. Examples follow.
Example 1.10 Let Ω contain all outcomes of a throw of a pair of six-sided
dice, and let P assign 1/36 to each outcome. Then Ω is the following set of
ordered pairs:
Ω = {(1, 1), (1, 2), (1, 3), (1, 4), (1, 5), (1, 6), (2, 1), (2, 2), . . . (6, 5), (6, 6)}.
Let the random variable
let the random variable
to a pair if at least one
table shows some of the
X assign the sum of each ordered pair to that pair, and
Y assign ‘odd’ to each pair of odd numbers and ‘even’
number in that pair is an even number. The following
values of X and Y :
e
(1, 1)
(1, 2)
···
(2, 1)
···
(6, 6)
X(e)
2
3
···
3
···
12
Y (e)
odd
even
···
even
···
even
The space of X is {2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12}, and that of Y is {odd, even}.
For a random variable X, we use X = x to denote the set of all elements
e ∈ Ω that X maps to the value of x. That is,
X =x
represents the event
{e such that X(e) = x}.
Note the difference between X and x. Small x denotes any element in the space
of X, while X is a function.
Example 1.11 Let Ω , P , and X be as in Example 1.10. Then
X=3
represents the event
P (X = 3) =
{(1, 2), (2, 1)} and
1
.
18
It is not hard to see that a random variable induces a probability function
on its space. That is, if we define PX ({x}) ≡ P (X = x), then PX is such a
probability function.
Example 1.12 Let Ω contain all outcomes of a throw of a single die, let P
assign 1/6 to each outcome, and let Z assign ‘even’ to each even number and
‘odd’ to each odd number. Then
1
PZ ({even}) = P (Z = even) = P ({2, 4, 6}) =
2
1.1. BASICS OF PROBABILITY THEORY
15
1
PZ ({odd}) = P (Z = odd) = P ({1, 3, 5}) = .
2
We rarely refer to PX ({x}). Rather we only reference the original probability
function P , and we call P (X = x) the probability distribution of the random
variable X. For brevity, we often just say ‘distribution’ instead of ‘probability
distribution’. Furthermore, we often use x alone to represent the event X = x,
and so we write P (x) instead of P (X = x) . We refer to P (x) as ‘the probability
of x’.
Let Ω, P , and X be as in Example 1.10. Then if x = 3,
1
.
18
Given two random variables X and Y , defined on the same sample space Ω,
we use X = x, Y = y to denote the set of all elements e ∈ Ω that are mapped
both by X to x and by Y to y. That is,
P (x) = P (X = x) =
X = x, Y = y
represents the event
{e such that X(e) = x} ∩ {e such that Y (e) = y}.
Example 1.13 Let Ω, P , X, and Y be as in Example 1.10. Then
X = 4, Y = odd
represents the event
{(1, 3), (3, 1)}, and
P (X = 4, Y = odd) = 1/18.
Clearly, two random variables induce a probability function on the Cartesian
product of their spaces. As is the case for a single random variable, we rarely
refer to this probability function. Rather we reference the original probability
function. That is, we refer to P (X = x, Y = y), and we call this the joint
probability distribution of X and Y . If A = {X, Y }, we also call this the
joint probability distribution of A. Furthermore, we often just say ‘joint
distribution’ or ‘probability distribution’.
For brevity, we often use x, y to represent the event X = x, Y = y, and
so we write P (x, y) instead of P (X = x, Y = y). This concept extends in a
straightforward way to three or more random variables. For example, P (X =
x, Y = y, Z = z) is the joint probability distribution function of the variables
X, Y , and Z, and we often write P (x, y, z).
Example 1.14 Let Ω, P , X, and Y be as in Example 1.10. Then if x = 4 and
y = odd,
P (x, y) = P (X = x, Y = y) = 1/18.
If, for example, we let A = {X, Y } and a = {x, y}, we use
A=a
to represent
X = x, Y = y,
and we often write P (a) instead of P (A = a). The same notation extends to
the representation of three or more random variables. For consistency, we set
P (∅ = ∅) = 1, where ∅ is the empty set of random variables. Note that if ∅
is the empty set of events, P (∅) = 0.
16
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
Example 1.15 Let Ω, P , X, and Y be as in Example 1.10. If A = {X, Y },
a = {x, y}, x = 4, and y = odd,
P (A = a) = P (X = x, Y = y) = 1/18.
This notation entails that if we have, for example, two sets of random variables A = {X, Y } and B = {Z, W }, then
A = a, B = b
represents
X = x, Y = y, Z = z, W = w.
Given a joint probability distribution, the law of total probability (Equality
1.1) implies the probability distribution of any one of the random variables
can be obtained by summing over all values of the other variables. It is left
as an exercise to show this. For example, suppose we have a joint probability
distribution P (X = x, Y = y). Then
P (X = x) =
X
P (X = x, Y = y),
y
P
where y means the sum as y goes through all values of Y . The probability
distribution P (X = x) is called the marginal probability distribution of X
because it is obtained using a process similar to adding across a row or column in
a table of numbers. This concept also extends in a straightforward way to three
or more random variables. For example, if we have a joint distribution P (X =
x, Y = y, Z = z) of X, Y , and Z, the marginal distribution P (X = x, Y = y) of
X and Y is obtained by summing over all values of Z. If A = {X, Y }, we also
call this the marginal probability distribution of A.
Example 1.16 Let Ω, P , X, and Y be as in Example 1.10. Then
P (X = 4) =
X
P (X = 4, Y = y)
y
= P (X = 4, Y = odd) + P (X = 4, Y = even) =
1
1
1
+
=
.
18 36
12
The following example reviews the concepts covered so far concerning random variables:
Example 1.17 Let Ω be a set of 12 individuals, and let P assign 1/12 to each
individual. Suppose the sexes, heights, and wages of the individuals are as follows:
1.1. BASICS OF PROBABILITY THEORY
Case
1
2
3
4
5
6
7
8
9
10
11
12
Sex
female
female
female
female
female
female
male
male
male
male
male
male
Height (inches)
64
64
64
64
68
68
64
64
68
68
70
70
17
Wage ($)
30, 000
30, 000
40, 000
40, 000
30, 000
40, 000
40, 000
50, 000
40, 000
50, 000
40, 000
50, 000
Let the random variables S, H and W respectively assign the sex, height and
wage of an individual to that individual. Then the distributions of the three
variables are as follows (Recall that, for example, P (s) represents P (S = s).):
s
female
male
P (s)
1/2
1/2
h
64
68
70
P (h)
1/2
1/3
1/6
w
30, 000
40, 000
50, 000
P (w)
1/4
1/2
1/4
The joint distribution of S and H is as follows:
s
female
female
female
male
male
male
h
64
68
70
64
68
70
P (s, h)
1/3
1/6
0
1/6
1/6
1/6
The following table also shows the joint distribution of S and H and illustrates
that the individual distributions can be obtained by summing the joint distribution over all values of the other variable:
64
68
70
Distribution of S
s
female
male
1/3
1/6
1/6
1/6
0
1/6
1/2
1/2
Distribution of H
1/2
1/3
1/6
h
The table that follows shows the first few values in the joint distribution of S,
H, and W . There are 18 values in all, of which many are 0.
18
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
s
female
female
female
female
···
h
64
64
64
68
···
w
30, 000
40, 000
50, 000
30, 000
···
P (s, h, w)
1/6
1/6
0
1/12
···
We have the following definition:
Definition 1.6 Suppose we have a probability space (Ω, P ), and two sets A and
B containing random variables defined on Ω. Then the sets A and B are said to
be independent if, for all values of the variables in the sets a and b, the events
A = a and B = b are independent. That is, either P (a) = 0 or P (b) = 0 or
P (a|b) = P (a).
When this is the case, we write
IP (A, B),
where IP stands for independent in P .
Example 1.18 Let Ω be the set of all cards in an ordinary deck, and let P
assign 1/52 to each card. Define random variables as follows:
Variable
R
T
S
Value
r1
r2
t1
t2
s1
s2
Outcomes Mapped to this Value
All royal cards
All nonroyal cards
All tens and jacks
All cards that are neither tens nor jacks
All spades
All nonspades
Then we maintain the sets {R, T } and {S} are independent. That is,
IP ({R, T }, {S}).
To show this, we need show for all values of r, t, and s that
P (r, t|s) = P (r, t).
(Note that it we do not show brackets to denote sets in our probabilistic expression because in such an expression a set represents the members of the set. See
the discussion following Example 1.14.) The following table shows this is the
case:
1.1. BASICS OF PROBABILITY THEORY
s
s1
s1
s1
s1
s2
s2
s2
s2
r
r1
r1
r2
r2
r1
r1
r2
r2
t
t1
t2
t1
t2
t1
t2
t1
t2
P (r, t|s)
1/13
2/13
1/13
9/13
3/39 = 1/13
6/39 = 2/13
3/39 = 1/13
27/39 = 9/13
19
P (r, t)
4/52 = 1/13
8/52 = 2/13
4/52 = 1/13
36/52 = 9/13
4/52 = 1/13
8/52 = 2/13
4/52 = 1/13
36/52 = 9/13
Definition 1.7 Suppose we have a probability space (Ω, P ), and three sets A,
B, and C containing random variable defined on Ω. Then the sets A and B are
said to be conditionally independent given the set C if, for all values of
the variables in the sets a, b, and c, whenever P (c) 6= 0, the events A = a and
B = b are conditionally independent given the event C = c. That is, either
P (a|c) = 0 or P (b|c) = 0 or
P (a|b, c) = P (a|c).
When this is the case, we write
IP (A, B|C).
Example 1.19 Let Ω be the set of all objects in Figure 1.2, and let P assign
1/13 to each object. Define random variables S (for shape), V (for value), and
C (for color) as follows:
Variable
V
S
C
Value
v1
v2
s1
s2
c1
c2
Outcomes Mapped to this Value
All objects containing a ‘1’
All objects containing a ‘2’
All square objects
All round objects
All black objects
All white objects
Then we maintain that {V } and {S} are conditionally independent given {C}.
That is,
IP ({V }, {S}|{C}).
To show this, we need show for all values of v, s, and c that
P (v|s, c) = P (v|c).
The results in Example 1.8 show P (v1|s1, c1) = P (v1|c1) and P (v1|s1, c2) =
P (v1|c2). The table that follows shows the equality holds for the other values of
the variables too:
20
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
c
c1
c1
c1
c1
c2
c2
c2
c2
s
s1
s1
s2
s2
s1
s1
s2
s2
v
v1
v2
v1
v2
v1
v2
v1
v2
P (v|s, c)
2/6 = 1/3
4/6 = 2/3
1/3
2/3
1/2
1/2
1/2
1/2
P (v|c)
3/9 = 1/3
6/9 = 2/3
3/9 = 1/3
6/9 = 2/3
2/4 = 1/2
2/4 = 1/2
2/4 = 1/2
2/4 = 1/2
For the sake of brevity, we sometimes only say ‘independent’ rather than
‘conditionally independent’. Furthermore, when a set contains only one item,
we often drop the set notation and terminology. For example, in the preceding
example, we might say V and S are independent given C and write IP (V, S|C).
Finally, we have the chain rule for random variables, which says that given
n random variables X1 , X2 , . . . Xn , defined on the same sample space Ω,
regola della catena
P (x1 , x2 , . . .xn ) = P (xn |xn−1 , xn−2 , . . .x1 ) · · · P (x2 |x1 )P (x1 )
whenever P (x1 , x2 , . . .xn ) 6= 0. It is straightforward to prove this rule using the
rule for conditional probability.
1.2
Bayesian Inference
We use Bayes’ Theorem when we are not able to determine the conditional
probability of interest directly, but we are able to determine the probabilities
on the right in Equality 1.3. You may wonder why we wouldn’t be able to
compute the conditional probability of interest directly from the sample space.
The reason is that in these applications the probability space is not usually
developed in the order outlined in Section 1.1. That is, we do not identify a
sample space, determine probabilities of elementary events, determine random
variables, and then compute values in joint probability distributions. Instead, we
identify random variables directly, and we determine probabilistic relationships
among the random variables. The conditional probabilities of interest are often
not the ones we are able to judge directly. We discuss next the meaning of
random variables and probabilities in Bayesian applications and how they are
identified directly. After that, we show how a joint probability distribution can
be determined without first specifying a sample space. Finally, we show a useful
application of Bayes’ Theorem.
1.2.1
Random Variables and Probabilities in Bayesian Applications
Although the definition of a random variable (Definition 1.5) given in Section
1.1.4 is mathematically elegant and in theory pertains to all applications of
probability, it is not readily apparent how it applies to applications involving
1.2. BAYESIAN INFERENCE
21
Bayesian inference. In this subsection and the next we develop an alternative
definition that does.
When doing Bayesian inference, there is some entity which has features,
the states of which we wish to determine, but which we cannot determine for
certain. So we settle for determining how likely it is that a particular feature is
in a particular state. The entity might be a single system or a set of systems.
An example of a single system is the introduction of an economically beneficial
chemical which might be carcinogenic. We would want to determine the relative
risk of the chemical versus its benefits. An example of a set of entities is a set
of patients with similar diseases and symptoms. In this case, we would want to
diagnose diseases based on symptoms.
In these applications, a random variable represents some feature of the entity
being modeled, and we are uncertain as to the values of this feature for the
particular entity. So we develop probabilistic relationships among the variables.
When there is a set of entities, we assume the entities in the set all have the same
probabilistic relationships concerning the variables used in the model. When
this is not the case, our Bayesian analysis is not applicable. In the case of the
chemical introduction, features may include the amount of human exposure and
the carcinogenic potential. If these are our features of interest, we identify the
random variables HumanExposure and CarcinogenicP otential (For simplicity,
our illustrations include only a few variables. An actual application ordinarily
includes many more than this.). In the case of a set of patients, features of
interest might include whether or not a disease such as lung cancer is present,
whether or not manifestations of diseases such as a chest X-ray are present,
and whether or not causes of diseases such as smoking are present. Given these
features, we would identify the random variables ChestXray, LungCancer,
and SmokingHistory. After identifying the random variables, we distinguish a
set of mutually exclusive and exhaustive values for each of them. The possible
values of a random variable are the different states that the feature can take.
For example, the state of LungCancer could be present or absent, the state of
ChestXray could be positive or negative, and the state of SmokingHistory
could be yes or no. For simplicity, we have only distinguished two possible
values for each of these random variables. However, in general they could have
any number of possible values or they could even be continuous. For example,
we might distinguish 5 different levels of smoking history (one pack or more
for at least 10 years, two packs or more for at least 10 years, three packs or
more for at lest ten years, etc.). The specification of the random variables and
their values not only must be precise enough to satisfy the requirements of the
particular situation being modeled, but it also must be sufficiently precise to
pass the clarity test, which was developed by Howard in 1988. That test
is as follows: Imagine a clairvoyant who knows precisely the current state of
the world (or future state if the model concerns events in the future). Would
the clairvoyant be able to determine unequivocally the value of the random
variable? For example, in the case of the chemical introduction, if we give
HumanExposure the values low and high, the clarity test is not passed because
we do not know what constitutes high or low. However, if we define high as
22
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
when the average (over all individuals), of the individual daily average skin
contact, exceeds 6 grams of material, the clarity test is passed because the
clairvoyant can answer precisely whether the contact exceeds that. In the case
of a medical application, if we give SmokingHistory only the values yes and
no, the clarity test is not passed because we do not know whether yes means
smoking cigarettes, cigars, or something else, and we have not specified how
long smoking must have occurred for the value to be yes. On the other hand, if
we say yes means the patient has smoked one or more packs of cigarettes every
day during the past 10 years, the clarity test is passed.
After distinguishing the possible values of the random variables (i.e. their
spaces), we judge the probabilities of the random variables having their values.
However, in general we do not always determine prior probabilities; nor do we determine values in a joint probability distribution of the random variables. Rather
we ascertain probabilities, concerning relationships among random variables,
that are accessible to us. For example, we might determine the prior probability
P (LungCancer = present), and the conditional probabilities P (ChestXray =
positive|LungCancer = present), P (ChestXray = positive|LungCancer =
absent), P (LungCancer = present| SmokingHistory = yes), and finally
P (LungCancer = present|SmokingHistory = no). We would obtain these
probabilities either from a physician or from data or from both. Thinking in
terms of relative frequencies, P (LungCancer = present|SmokingHistory =
yes) can be estimated by observing individuals with a smoking history, and determining what fraction of these have lung cancer. A physician is used to judging
such a probability by observing patients with a smoking history. On the other
hand, one does not readily judge values in a joint probability distribution such as
P (LungCancer = present, ChestXray = positive, SmokingHistory = yes). If
this is not apparent, just think of the situation in which there are 100 or more
random variables (which there are in some applications) in the joint probability
distribution. We can obtain data and think in terms of probabilistic relationships among a few random variables at a time; we do not identify the joint
probabilities of several events.
As to the nature of these probabilities, consider first the introduction of the
toxic chemical. The probabilities of the values of CarcinogenicP otential will
be based on data involving this chemical and similar ones. However, this is
certainly not a repeatable experiment like a coin toss, and therefore the probabilities are not relative frequencies. They are subjective probabilities based
on a careful analysis of the situation. As to the medical application involving a set of entities, we often obtain the probabilities from estimates of relative frequencies involving entities in the set. For example, we might obtain
P (ChestXray = positive|LungCancer = present) by observing 1000 patients
with lung cancer and determining what fraction have positive chest X-rays.
However, as will be illustrated in Section 1.2.3, when we do Bayesian inference
using these probabilities, we are computing the probability of a specific individual being in some state, which means it is a subjective probability. Recall from
Section 1.1.1 that a relative frequency is not a property of any one of the trials
(patients), but rather it is a property of the entire sequence of trials. You may
1.2. BAYESIAN INFERENCE
23
feel that we are splitting hairs. Namely, you may argue the following: “This
subjective probability regarding a specific patient is obtained from a relative
frequency and therefore has the same value as it. We are simply calling it a
subjective probability rather than a relative frequency.” But even this is not
the case. Even if the probabilities used to do Bayesian inference are obtained
from frequency data, they are only estimates of the actual relative frequencies.
So they are subjective probabilities obtained from estimates of relative frequencies; they are not relative frequencies. When we manipulate them using Bayes’
theorem, the resultant probability is therefore also only a subjective probability.
Once we judge the probabilities for a given application, we can often obtain values in a joint probability distribution of the random variables. Theorem 1.5 in Section 1.3.3 obtains a way to do this when there are many variables. Presently, we illustrate the case of two variables. Suppose we only
identify the random variables LungCancer and ChestXray, and we judge the
prior probability P (LungCancer = present), and the conditional probabilities P (ChestXray = positive|LungCancer = present) and P (ChestXray =
positive|LungCancer = absent). Probabilities of values in a joint probability
distribution can be obtained from these probabilities using the rule for conditional probability as follows:
P (present, positive) = P (positive|present)P (present)
P (present, negative) = P (negative|present)P (present)
P (absent, positive) = P (positive|absent)P (absent)
P (absent, negative) = P (negative|absent)P (absent).
Note that we used our abbreviated notation. We see then that at the outset we
identify random variables and their probabilistic relationships, and values in a
joint probability distribution can then often be obtained from the probabilities
relating the random variables. So what is the sample space? We can think of the
sample space as simply being the Cartesian product of the sets of all possible
values of the random variables. For example, consider again the case where we
only identify the random variables LungCancer and ChestXray, and ascertain
probability values in a joint distribution as illustrated above. We can define the
following sample space:
Ω=
{(present, positive), (present, negative), (absent, positive), (absent, negative)}.
We can consider each random variable a function on this space that maps
each tuple into the value of the random variable in the tuple. For example,
LungCancer would map (present, positive) and (present, negative) each into
present. We then assign each elementary event the probability of its corresponding event in the joint distribution. For example, we assign
P̂ ({(present, positive)}) = P (LungCancer = present, ChestXray = positive).
24
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
It is not hard to show that this does yield a probability function on Ω and
that the initially assessed prior probabilities and conditional probabilities are
the probabilities they notationally represent in this probability space (This is a
special case of Theorem 1.5.).
Since random variables are actually identified first and only implicitly become functions on an implicit sample space, it seems we could develop the concept of a joint probability distribution without the explicit notion of a sample
space. Indeed, we do this next. Following this development, we give a theorem
showing that any such joint probability distribution is a joint probability distribution of the random variables with the variables considered as functions on
an implicit sample space. Definition 1.1 (of a probability function) and Definition 1.5 (of a random variable) can therefore be considered the fundamental
definitions for probability theory because they pertains both to applications
where sample spaces are directly identified and ones where random variables
are directly identified.
1.2.2
A Definition of Random Variables and Joint Probability Distributions for Bayesian Inference
For the purpose of modeling the types of problems discussed in the previous
subsection, we can define a random variable X as a symbol representing any
one of a set of values, called the space of X. For simplicity, we will assume
the space of X is countable, but the theory extends naturally to the case where
it is not. For example, we could identify the random variable LungCancer as
having the space {present, absent}. We use the notation X = x as a primitive
which is used in probability expressions. That is, X = x is not defined in terms
of anything else. For example, in application LungCancer = present means the
entity being modeled has lung cancer, but mathematically it is simply a primitive which is used in probability expressions. Given this definition and primitive,
we have the following direct definition of a joint probability distribution:
Definition 1.8 Let a set of n random variables V = {X1 , X2 , . . . Xn } be specified such that each Xi has a countably infinite space. A function, that assigns a
real number P (X1 = x1 , X2 = x2 , . . . Xn = xn ) to every combination of values
of the xi ’s such that the value of xi is chosen from the space of Xi , is called a
joint probability distribution of the random variables in V if it satisfies the
following conditions:
1. For every combination of values of the xi ’s,
0 ≤ P (X1 = x1 , X2 = x2 , . . . Xn = xn ) ≤ 1.
2. We have
X
x1 ,x2,... xn
P (X1 = x1 , X2 = x2 , . . . Xn = xn ) = 1.
1.2. BAYESIAN INFERENCE
25
P
The notation
x1 ,x2,... xn means the sum as the variables x1 , . . . xn go
through all possible values in their corresponding spaces.
Note that a joint probability distribution, obtained by defining random variables as functions on a sample space, is one way to create a joint probability
distribution that satisfies this definition. However, there are other ways as the
following example illustrates:
Example 1.20 Let V = {X, Y }, let X and Y have spaces {x1, x2}1 and {y1, y2}
respectively, and let the following values be specified:
P (X = x1) = .2
P (X = x2) = .8
P (Y = y1) = .3
P (Y = y2) = .7.
Next define a joint probability distribution of X and Y as follows:
P (X = x1, Y = y1) = P (X = x1)P (Y = y1) = (.2)(.3) = .06
P (X = x1, Y = y2) = P (X = x1)P (Y = y2) = (.2)(.7) = .14
P (X = x2, Y = y1) = P (X = x2)P (Y = y1) = (.8)(.3) = .24
P (X = x2, Y = y2) = P (X = x2)P (Y = y2) = (.8)(.7) = .56.
Since the values sum to 1, this is another way of specifying a joint probability
distribution according to Definition 1.8. This is how we would specify the joint
distribution if we felt X and Y were independent.
Notice that our original specifications, P (X = xi) and P (Y = yi), notationally look like marginal distributions of the joint distribution developed in
Example 1.20. However, Definition 1.8 only defines a joint probability distribution P ; it does not mention anything about marginal distributions. So the
initially specified values do not represent marginal distributions of our joint distribution P according to that definition alone. The following theorem enables
us to consider them marginal distributions in the classical sense, and therefore
justifies our notation.
Theorem 1.3 Let a set of random variables V be given and let a joint probability distribution of the variables in V be specified according to Definition 1.8.
Let Ω be the Cartesian product of the sets of all possible values of the random
variables. Assign probabilities to elementary events in Ω as follows:
P̂ ({(x1 , x2 , . . . xn )}) = P (X1 = x1 , X2 = x2 , . . . Xn = xn ).
These assignments result in a probability function on Ω according to Definition
1.1. Furthermore, if we let X̂i denote a function (random variable in the classical sense) on this sample space that maps each tuple in Ω to the value of xi in
1 We use subscripted variables X to denote different random variables. So we do not
i
subcript to denote a value of a random variable. Rather we write the index next to the
variable.
26
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
that tuple, then the joint probability distribution of the X̂i ’s is the same as the
originally specified joint probability distribution.
Proof. The proof is left as an exercise.
Example 1.21 Suppose we directly specify a joint probability distribution of X
and Y , each with space {x1, x2} and {y1, y2} respectively, as done in Example
1.20. That is, we specify the following probabilities:
P (X
P (X
P (X
P (X
= x1, Y
= x1, Y
= x2, Y
= x2, Y
= y1)
= y2)
= y1)
= y2).
Next we let Ω = {(x1, y1), (x1, y2), (x2, y1), (x2, y2)}, and we assign
P̂ ({(xi, yj)}) = P (X = xi, Y = yj).
Then we let X̂ and Ŷ be functions on Ω defined by the following tables:
x
x1
x1
x2
x2
y
y1
y2
y1
y2
X̂((x, y))
x1
x1
x2
x2
x
x1
x1
x2
x2
y
y1
y2
y1
y2
Ŷ ((x, y))
y1
y2
y1
y2
Theorem 1.3 says the joint probability distribution of these random variables is
the same as the originally specified joint probability distribution. Let’s illustrate
this:
P̂ (X̂ = x1, Ŷ = y1) = P̂ ({(x1, y1), (x1, y2)} ∩ {(x1, y1), (x2, y1)})
= P̂ ({(x1, y1)})
= P (X = x1, Y = y1).
Due to Theorem 1.3, we need no postulates for probabilities of combinations
of primitives not addressed by Definition 1.8. Furthermore, we need no new
definition of conditional probability for joint distributions created according
to that definition. We can just postulate that both obtain values according
to the set theoretic definition of a random variable. For example, consider
Example 1.20. Due to Theorem 1.3, P̂ (X̂ = x1) is simply a value in a marginal
distribution of the joint probability distribution. So its value is computed as
follows:
P̂ (X̂ = x1) =
=
=
=
=
P̂ (X̂
P (X
P (X
P (X
P (X
= x1, Ŷ = y1) + P̂ (X̂ = x1, Ŷ = y2)
= x1, Y = y1) + P (X = x1, Y = y2)
= x1)P (Y = y1) + P (X = x1)P (Y = y2)
= x1)[P (Y = y1) + P (Y = y2)]
= x1)[1] = P (X = x1),
1.2. BAYESIAN INFERENCE
27
which is the originally specified value. This result is a special case of Theorem
1.5.
Note that the specified probability values are not by necessity equal to the
probabilities they notationally represent in the marginal probability distribution. However, since we used the rule for independence to derive the joint
probability distribution from them, they are in fact equal to those values. For
example, if we had defined P (X = x1, Y = y1) = P (X = x2)P (Y = y1), this
would not be the case. Of course we would not do this. In practice, all specified
values are always the probabilities they notationally represent in the resultant
probability space (Ω, P̂ ). Since this is the case, we will no longer show carats
over P or X when referring to the probability function in this space or a random
variable on the space.
Example 1.22 Let V = {X, Y }, let X and Y have spaces {x1, x2} and {y1, y2}
respectively, and let the following values be specified:
P (X = x1) = .2
P (X = x2) = .8
P (Y = y1|X = x1) = .3
P (Y = y2|X = x1) = .7
P (Y = y1|X = x2) = .4
P (Y = y2|X = x2) = .6.
Next define a joint probability distribution of X and Y as follows:
P (X = x1, Y = y1) = P (Y = y1|X = x1)P (X = x1) = (.3)(.2) = .06
P (X = x1, Y = y2) = P (Y = y2|X = x1)P (X = x1) = (.7)(.2) = .14
P (X = x2, Y = y1) = P (Y = y1|X = x2)P (X = x2) = (.4)(.8) = .32
P (X = x2, Y = y2) = P (Y = y2|X = x2)P (X = x2) = (.6)(.8) = .48.
Since the values sum to 1, this is another way of specifying a joint probability
distribution according to Definition 1.8. As we shall see in Example 1.23 in the
following subsection, this is the way they are specified in simple applications of
Bayes’ Theorem.
In the remainder of this text, we will create joint probability distributions
using Definition 1.8. Before closing, we note that this definition pertains to any
application in which we model naturally occurring phenomena by identifying
random variables directly, which includes most applications of statistics.
1.2.3
A Classical Example of Bayesian Inference
The following examples illustrates how Bayes’ theorem has traditionally been
applied to compute the probability of an event of interest from known probabilities.
28
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
Example 1.23 Suppose Joe has a routine diagnostic chest X-ray required of
all new employees at Colonial Bank, and the X-ray comes back positive for lung
cancer. Joe then becomes certain he has lung cancer and panics. But should
he? Without knowing the accuracy of the test, Joe really has no way of knowing
how probable it is that he has lung cancer. When he discovers the test is not
absolutely conclusive, he decides to investigate its accuracy and he learns that it
has a false negative rate of .4 and a false positive rate of .02. We represent this
accuracy as follows. First we define these random variables:
Variable
T est
LungCancer
Value
positive
negative
present
absent
When the Variable Takes This Value
X-ray is positive
X-ray is negative
Lung cancer is present
Lung cancer is absent
We then have these conditional probabilities:
P (T est = positive|LungCancer = present) = .6
P (T est = positive|LungCancer = absent) = .02.
Given these probabilities, Joe feels a little better. However, he then realizes he
still does not know how probable it is that he has lung cancer. That is, the probability of Joe having lung cancer is P (LungCancer = present|T est = positive),
and this is not one of the probabilities listed above. Joe finally recalls Bayes’
theorem and realizes he needs yet another probability to determine the probability
of his having lung cancer. That probability is P (LungCancer = present), which
is the probability of his having lung cancer before any information on the test
results were obtained. Even though this probability is not based on any information concerning the test results, it is based on some information. Specifically, it
is based on all information (relevant to lung cancer) known about Joe before he
took the test. The only information about Joe, before he took the test, was that
he was one of a class of employees who took the test routinely required of new
employees. So, when he learns only 1 out of every 1000 new employees has lung
cancer, he assigns .001 to P (LungCancer = present). He then employs Bayes’
theorem as follows (Note that we again use our abbreviated notation):
P (present|positive)
P (positive|present)P (present)
P (positive|present)P (present) + P (positive|absent)P (absent)
(.6)(.001)
=
(.6)(.001) + (.02)(.999)
= .029.
=
So Joe now feels that he probability of his having lung cancer is only about .03,
and he relaxes a bit while waiting for the results of further testing.
1.3. LARGE INSTANCES / BAYESIAN NETWORKS
29
A probability like P (LungCancer = present) is called a prior probability
because, in a particular model, it is the probability of some event prior to
updating the probability of that event, within the framework of that model,
using new information. Do not mistakenly think it means a probability prior to
any information. A probability like P (LungCancer = present|T est = positive)
is called a posterior probability because it is the probability of an event
after its prior probability has been updated, within the framework of some
model, based on new information. The following example illustrates how prior
probabilities can change depending on the situation we are modeling.
Example 1.24 Now suppose Sam is having the same diagnostic chest X-ray
as Joe. However, he is having the X-ray because he has worked in the mines
for 20 years, and his employers became concerned when they learned that about
10% of all such workers develop lung cancer after many years in the mines.
Sam also tests positive. What is the probability he has lung cancer? Based on
the information known about Sam before he took the test, we assign a prior
probability of .1 to Sam having lung cancer. Again using Bayes’ theorem, we
conclude that P (LungCancer = present|T est = positive) = .769 for Sam. Poor
Sam concludes it is quite likely that he has lung cancer.
The previous two examples illustrate that a probability value is relative to
one’s information about an event; it is not a property of the event itself. Both
Joe and Sam either do or do not have lung cancer. It could be that Joe has
it and Sam does not. However, based on our information, our degree of belief
(probability) that Sam has it is much greater than our degree of belief that Joe
has it. When we obtain more information relative to the event (e.g. whether
Joe smokes or has a family history of cancer), the probability will change.
1.3
Large Instances / Bayesian Networks
Bayesian inference is fairly simple when it involves only two related variables as
in Example 1.23. However, it becomes much more complex when we want to
do inference with many related variable. We address this problem next. After
discussing the difficulties inherent in representing large instances and in doing
inference when there are a large number of variables, we describe a relationship, called the Markov condition, between graphs and probability distributions.
Then we introduce Bayesian networks, which exploit the Markov condition in
order to represent large instances efficiently.
1.3.1
The Difficulties Inherent in Large Instances
Recall the situation, discussed at the beginning of this chapter, where several
features (variables) are related through inference chains. We introduced the
following example of this situation: Whether or not an individual has a history
of smoking has a direct influence both on whether or not that individual has
bronchitis and on whether or not that individual has lung cancer. In turn, the
30
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
presence or absence of each of these features has a direct influence on whether
or not the individual experiences fatigue. Also, the presence or absence of
lung cancer has a direct influence on whether or not a chest X-ray is positive.
We noted that, in this situation, we would want to do probabilistic inference
involving features that are not related via a direct influence. We would want to
determine, for example, the conditional probabilities both of having bronchitis
and of having lung cancer when it is known an individual smokes, is fatigued,
and has a positive chest X-ray. Yet bronchitis has no influence on whether a
chest X-ray is positive. Therefore, this conditional probability cannot readily
be computed using a simple application of Bayes’ theorem. So how could we
compute it? Next we develop a straightforward algorithm for doing so, but we
will show it has little practical value. First we give some notation. As done
previously, we will denote random variables using capital letters such as X and
use the corresponding lower case letters x1, x2, etc. to denote the values in the
space of X. In the current example, we define the random variables that follow:
Variable
H
B
L
F
C
Value
h1
h2
b1
b2
l1
l2
f1
f2
c1
c2
When the Variable Takes this Value
There is a history of smoking
There is no history of smoking
Bronchitis is present
Bronchitis is absent
Lung cancer is present
Lung cancer is absent
Fatigue is present
Fatigue is absent
Chest X-ray is positive
Chest X-ray is negative
Note that we presented this same table at the beginning of this chapter, but we
called the random variables ‘features’. We had not yet defined random variable
at that point; so we used the informal term feature. If we knew the joint
probability distribution of these five variables, we could compute the conditional
probability of an individual having bronchitis given the individual smokes, is
fatigued, and has a positive chest X-ray as follows:
P
P (b1, h1, f1, c1, l)
P (b1, h1, f 1, c1)
l
= P
,
(1.5)
P (b1|h1, f 1, c1) =
P (h1, f1, c1)
P (b, h1, f 1, c1, l)
b,l
P
where b,l means the sum as b and l go through all their possible values. There
are a number of problems here. First, as noted previously, the values in the joint
probability distribution are ordinarily not readily accessible. Second, there are
an exponential number of terms in the sums in Equality 1.5. That is, there
are 22 terms in the sum in the denominator, and, if there were 100 variables
in the application, there would be 297 terms in that sum. So, in the case
of a large instance, even if we had some means for eliciting the values in the
1.3. LARGE INSTANCES / BAYESIAN NETWORKS
31
joint probability distribution, using Equality 1.5 simply requires determining
too many such values and doing too many calculations with them. We see that
this method has no practical value when the instance is large.
Bayesian networks address the problems of 1) representing the joint probability distribution of a large number of random variables; and 2) doing Bayesian
inference with these variables. Before introducing them in Section 1.3.3, we
need to discuss the Markov condition.
1.3.2
The Markov Condition
First let’s review some graph theory. Recall that a directed graph is a pair
(V, E), where V is a finite, nonempty set whose elements are called nodes (or
vertices), and E is a set of ordered pairs of distinct elements of V. Elements of
E are called edges (or arcs), and if (X, Y ) ∈ E, we say there is an edge from
X to Y and that X and Y are each incident to the edge. If there is an edge
from X to Y or from Y to X, we say X and Y are adjacent. Suppose we have
a set of nodes [X1 , X2 , . . . Xk ], where k ≥ 2, such (Xi−1 , Xi ) ∈ E for 2 ≤ i ≤ k.
We call the set of edges connecting the k nodes a path from X1 to Xk . The
nodes X2 , . . . Xk−1 are called interior nodes on path [X1 , X2 , . . . Xk ]. The
subpath of path [X1 , X2 , . . . Xk ] from Xi to Xj is the path [Xi , Xi+1 , . . . Xj ]
where 1 ≤ i < j ≤ k. A directed cycle is a path from a node to itself. A
simple path is a path containing no subpaths which are directed cycles. A
directed graph G is called a directed acyclic graph (DAG) if it contains no
directed cycles. Given a DAG G = (V, E) and nodes X and Y in V, Y is called
a parent of X if there is an edge from Y to X, Y is called a descendent of
X and X is called an ancestor of Y if there is a path from X to Y , and Y is
called a nondescendent of X if Y is not a descendent of X. Note that in this
text X is not considered a descendent of X because we require k ≥ 2 in the
definition of a path. Some texts say there is an empty path from X to X.
We can now state the following definition:
Definition 1.9 Suppose we have a joint probability distribution P of the random variables in some set V and a DAG G = (V, E). We say that (G, P ) satisfies
the Markov condition if for each variable X ∈ V, {X} is conditionally independent of the set of all its nondescendents given the set of all its parents.
Using the notation established in Section 1.1.4, this means if we denote the sets
of parents and nondescendents of X by PAX and NDX respectively, then
IP ({X}, NDX |PAX ).
When (G, P ) satisfies the Markov condition, we say G and P satisfy the
Markov condition with each other.
If X is a root, then its parent set PAX is empty. So in this case the Markov
condition means {X} is independent of NDX . That is, IP ({X}, NDX ). It is
not hard to show that IP ({X}, NDX |PAX ) implies IP ({X}, B|PAX ) for any
B ⊆ NDX . It is left as an exercise to do this. Notice that PAX ⊆ NDX . So
we could define the Markov condition by saying that X must be conditionally
32
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
C
V
V
C
S
S
(a)
(b)
V
V
C
S
S
C
(c)
(d)
Figure 1.3: The probability distribution in Example 1.25 satisfies the Markov
condition only for the DAGs in (a), (b), and (c).
independent of NDX − PAX given PAX . However, it is standard to define it as
above. When discussing the Markov condition relative to a particular distribution and DAG (as in the following examples), we just show the conditional
independence of X and NDX − PAX .
Example 1.25 Let Ω be the set of objects in Figure 1.2, and let P assign a
probability of 1/13 to each object. Let random variables V , S, and C be as
defined as in Example 1.19. That is, they are defined as follows:
Variable
V
S
C
Value
v1
v2
s1
s2
c1
c2
Outcomes Mapped to this Value
All objects containing a ‘1’
All objects containing a ‘2’
All square objects
All round objects
All black objects
All white objects
1.3. LARGE INSTANCES / BAYESIAN NETWORKS
33
H
B
L
F
C
Figure 1.4: A DAG illustrating the Markov condition
Then, as shown in Example 1.19, IP ({V }, {S}|{C}). Therefore, (G, P ) satisfies
the Markov condition if G is the DAG in Figure 1.3 (a), (b), or (c). However,
(G, P ) does not satisfy the Markov condition if G is the DAG in Figure 1.3 (d)
because IP ({V }, {S}) is not the case.
Example 1.26 Consider the DAG G in Figure 1.4. If (G, P ) satisfied the
Markov condition for some probability distribution P , we would have the following conditional independencies:
Node
C
B
F
L
PA
{L}
{H}
{B, L}
{H}
Conditional Independency
IP ({C}, {H, B, F }|{L})
IP ({B}, {L, C}|{H})
IP ({F }, {H, C}|{B, L})
IP ({L}, {B}|{H})
Recall from Section 1.3.1 that the number of terms in a joint probability
distribution is exponential in terms of the number of variables. So, in the
case of a large instance, we could not fully describe the joint distribution by
determining each of its values directly. Herein lies one of the powers of the
Markov condition. Theorem 1.4, which follows shortly, shows if (G, P ) satisfies
the Markov condition, then P equals the product of its conditional probability
distributions of all nodes given values of their parents in G, whenever these
conditional distributions exist. After proving this theorem, we discuss how this
means we often need ascertain far fewer values than if we had to determine all
values in the joint distribution directly. Before proving it, we illustrate what it
means for a joint distribution to equal the product of its conditional distributions
of all nodes given values of their parents in a DAG G. This would be the case
34
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
for a joint probability distribution P of the variables in the DAG in Figure 1.4
if, for all values of f , c, b, l, and h,
P (f, c, b, l, h) = P (f |b, l)P (c|l)P (b|h)P (l|h)P (h),
(1.6)
whenever the conditional probabilities on the right exist. Notice that if one of
them does not exist for some combination of the values of the variables, then
P (b, l) = 0 or P (l) = 0 or P (h) = 0, which implies P (f, c, b, l, h) = 0 for that
combination of values. However, there are cases in which P (f, c, b, l, h) = 0 and
the conditional probabilities still exist. For example, this would be the case if
all the conditional probabilities on the right existed and P (f|b, l) = 0 for some
combination of values of f , b, and l. So Equality 1.6 must hold for all nonzero
values of the joint probability distribution plus some zero values.
We now give the theorem.
Theorem 1.4 If (G, P ) satisfies the Markov condition, then P is equal to the
product of its conditional distributions of all nodes given values of their parents,
whenever these conditional distributions exist.
Proof. We prove the case where P is discrete. Order the nodes so that if Y is
a descendent of Z, then Y follows Z in the ordering. Such an ordering is called
an ancestral ordering. Examples of such an ordering for the DAG in Figure
1.4 are [H, L, B, C, F ] and [H, B, L, F, C]. Let X1 , X2 , . . . Xn be the resultant
ordering. For a given set of values of x1 , x2 , . . . xn , let pai be the subset of
these values containing the values of Xi ’s parents. We need show that whenever
P (pai ) 6= 0 for 1 ≤ i ≤ n,
P (xn , xn−1 , . . . x1 ) = P (xn |pan )P (xn−1 |pan−1 ) · · · P (x1 |pa1 ).
We show this using induction on the number of variables in the network. Assume, for some combination of values of the xi ’s, that P (pai ) 6= 0 for 1 ≤ i ≤ n.
induction base: Since PA1 is empty,
P (x1 ) = P (x1 |pa1 ).
induction hypothesis: Suppose for this combination of values of the xi ’s that
P (xi , xi−1 , . . . x1 ) = P (xi |pai )P (xi−1 |pai−1 ) · · · P (x1 |pa1 ).
induction step: We need show for this combination of values of the xi ’s that
P (xi+1 , xi , . . . x1 ) = P (xi+1 |pai+1 )P (xi |pai ) · · · P (x1 |pa1 ).
(1.7)
There are two cases:
Case 1: For this combination of values
P (xi , xi−1 , . . . x1 ) = 0.
(1.8)
1.3. LARGE INSTANCES / BAYESIAN NETWORKS
35
Clearly, Equality 1.8 implies
P (xi+1 , xi , . . . x1 ) = 0.
Furthermore, due to Equality 1.8 and the induction hypothesis, there is some k,
where 1 ≤ k ≤ i, such that P (xk |pak ) = 0. So Equality 1.7 holds.
Case 2: For this combination of values
P (xi , xi−1 , . . . x1 ) 6= 0.
In this case,
P (xi+1 , xi , . . . x1 ) = P (xi+1 |xi , . . . x1 )P (xi , . . . x1 )
= P (xi+1 |pai+1 )P (xi , . . . x1 )
= P (xi+1 |pai+1 )P (xi |pai ) · · · P (x1 |pa1 ).
The first equality is due to the rule for conditional probability, the second is due
to the Markov condition and the fact that X1 , . . . Xi are all nondescendents of
Xi+1 , and the last is due to the induction hypothesis.
Example 1.27 Recall that the joint probability distribution in Example 1.25
satisfies the Markov condition with the DAG in Figure 1.3 (a). Therefore, owing
to Theorem 1.4,
P (v, s, c) = P (v|c)P (s|c)p(c),
(1.9)
and we need only determine the conditional distributions on the right in Equality
1.9 to uniquely determine the values in the joint distribution. We illustrate that
this is the case for v1, s1, and c1:
P (v1, s1, c1) = P (One ∩ Square ∩ Black) =
2
13
P (v1|c1)P (s1|c1)P (c1) = P (One|Black) × P (Square|Black) × P (Black)
9
2
1 2
× ×
=
.
=
3 3 13
13
Figure 1.5 shows the DAG along with the conditional distributions.
The joint probability distribution in Example 1.25 also satisfies the Markov
condition with the DAGs in Figures 1.3 (b) and (c). Therefore, the probability
distribution in that example equals the product of the conditional distributions
for each of them. You should verify this directly.
If the DAG in Figure 1.3 (d) and some probability distribution P satisfied
the Markov condition, Theorem 1.4 would imply
P (v, s, c) = P (c|v, s)P (v)p(s).
Such a distribution is discussed in Exercise 1.20.
36
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
P(c1) = 9/13
P(c2) = 4/13
C
V
S
P(v1|c1) = 1/3
P(v2|c1) = 2/3
P(s1|c1) = 2/3
P(s2|c1) = 1/3
P(v1|c2) = 1/2
P(v2|c2) = 1/2
P(s1|c2) = 1/2
P(s2|c2) = 1/2
Figure 1.5: The probability distribution discussed in Example 1.27 is equal to
the product of these conditional distributions.
Theorem 1.4 often enables us to reduce the problem of determining a huge
number of probability values to that of determining relatively few. The number of values in the joint distribution is exponential in terms of the number of
variables. However, each of these values is uniquely determined by the conditional distributions (due to the theorem), and, if each node in the DAG does
not have too many children, there are not many values in these distributions.
For example, if each variable has two possible values and each node has at most
one parent, we would need to ascertain less than 2n probability values to determine the conditional distributions when the DAG contains n nodes. On the
other hand, we would need to ascertain 2n − 1 values to determine the joint
probability distribution directly. In general, if each variable has two possible
values and each node has at most k parents, we need to ascertain less than 2k n
values to determine the conditional distributions. So if k is not large, we have
a manageable number of values.
Something may seem amiss to you. Namely, in Example 1.25, we started
with an underlying sample space and probability function, specified some random variables, and showed that if P is the probability distribution of these
variables and G is the DAG in Figure 1.3 (a), then (P, G) satisfies the Markov
condition. We can therefore apply Theorem 1.4 to conclude we need only determine the conditional distributions of the variables for that DAG to find any
value in the joint distribution. We illustrated this in Example 1.27. However, as discussed in Section 1.2, in application we do not ordinarily specify
an underlying sample space and probability function from which we can compute conditional distributions. Rather we identify random variables and values
in conditional distributions directly. For example, in an application involving the diagnosis of lung cancer, we identify variables like SmokingHistory,
LungCancer, and ChestXray, and probabilities such as P (SmokingHistory =
1.3. LARGE INSTANCES / BAYESIAN NETWORKS
37
yes), P (LungCancer = present|SmokingHistory = yes), and P (ChestXray =
positive| LungCancer = present). How do we know the product of these conditional distributions is a joint distribution at all, much less one satisfying the
Markov condition with some DAG? Theorem 1.4 tells us only that if we start
with a joint distribution satisfying the Markov condition with some DAG, the
values in that joint distribution will be given by the product of the conditional distributions. However, we must work in reverse. We must start with
the conditional distributions and then be able to conclude the product of these
distributions is a joint distribution satisfying the Markov condition with some
DAG. The theorem that follows enables us to do just that.
Theorem 1.5 Let a DAG G be given in which each node is a random variable,
and let a discrete conditional probability distribution of each node given values of
its parents in G be specified. Then the product of these conditional distributions
yields a joint probability distribution P of the variables, and (G, P ) satisfies the
Markov condition.
Proof. Order the nodes according to an ancestral ordering. Let X1 , X2 , . . . Xn
be the resultant ordering. Next define
P (x1 , x2 , . . . xn ) = P (xn |pan )P (xn−1 |pan−1 ) · · · P (x2 |pa2 )P (x1 |pa1 ),
where PAi is the set of parents of Xi of in G and P (xi |pai ) is the specified
conditional probability distribution. First we show this does indeed yield a joint
probability distribution. Clearly, 0 ≤ P (x1 , x2 , . . .xn ) ≤ 1 for all values of the
variables. Therefore, to show we have a joint distribution, Definition 1.8 and
Theorem 1.3 imply we need only show that the sum of P (x1 , x2 , . . . xn ), as the
variables range through all their possible values, is equal to one. To that end,
XX
XX
...
P (x1 , x2 , . . .xn )
x1
=
x2
xn−1 xn
XX
x1
x2
···
XX
xn−1 xn
P (xn |pan )P (xn−1 |pan−1 ) · · · P (x2 |pa2 )P (x1 |pa1 )




#
"
X X
X X

· · ·
=
P (xn |pan ) P (xn−1 |pan−1 ) · · ·  P (x2 |pa2 ) P (x1 |pa1 )
x1
x2
=
x2
X X
x1
=
"
X
x1
x2

xn−1
xn


X
X X

· · ·
[1] P (xn−1 |pan−1 ) · · ·  P (x2 |pa2 ) P (x1 |pa1 )
=
x1

xn−1
#
[· · · 1 · · · ] P (x2 |pa2 ) P (x1 |pa1 )
[1] P (x1 |pa1 ) = 1.
It is left as an exercise to show that the specified conditional distributions are
the conditional distributions they notationally represent in the joint distribution.
38
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
Finally, we show the Markov condition is satisfied. To do this, we need show
for 1 ≤ k ≤ n that whenever P (pak ) 6= 0, if P (ndk |pak ) 6= 0 and P (xk |pak ) 6= 0
then P (xk |ndk , pak ) = P (xk |pak ), where NDk is the set of nondescendents of
Xk of in G. Since PAk ⊆ NDk , we need only show P (xk |ndk ) = P (xk |pak ).
First for a given k, order the nodes so that all and only nondescendents of
Xk precede Xk in the ordering. Note that this ordering depends on k, whereas
the ordering in the first part of the proof does not. Clearly then
NDk = {X1 , X2 , . . . Xk−1 }.
Let
Dk = {Xk+1 , Xk+2 , . . . Xn }.
X
means the sum as the variables in dk go through all
In what follows,
dk
their possible values. Furthermore, notation such as x̂k means the variable has
a particular value; notation such as n̂dk means all variables in the set have
particular values; and notation such as pan means some variables in the set
may not have particular values. We have that
P (x̂k |n̂dk ) =
=
P (x̂k , n̂dk )
P (n̂dk )
X
P (x̂1 , x̂2 , . . .x̂k , xk+1 , . . .xn )
d
k
X
P (x̂1 , x̂2 , . . .x̂k−1 , xk , . . .xn )
dk ∪{xk }
=
X
dk
X
P (xn |pan ) · · · P (xk+1 |pak+1 )P (x̂k |p̂ak ) · · · P (x̂1 |p̂a1 )
dk ∪{xk }
P (xn |pan ) · · · P (xk |pak )P (x̂k−1 |p̂ak−1 ) · · · P (x̂1 |p̂a1 )
P (x̂k |p̂ak ) · · · P (x̂1 |p̂a1 )
=
X
dk
P (x̂k−1 |p̂ak−1 ) · · · P (x̂1 |p̂a1 )
=
P (x̂k |p̂ak ) [1]
= P (x̂k |p̂ak ).
[1]
P (xn |pan ) · · · P (xk+1 |pak+1 )
X
dk ∪{xk }
P (xn |pan ) · · · P (xk |pak )
In the second to last step, the sums are each equal to one for the following reason.
Each is a sum of a product of conditional probability distributions specified for
a DAG. In the case of the numerator, that DAG is the subgraph, of our original
DAG G, consisting of the variables in Dk , and in the case of the denominator,
it is the subgraph consisting of the variables in Dk ∪{Xk }. Therefore, the fact
that each sum equals one follows from the first part of this proof.
Notice that the theorem requires that specified conditional distributions be
discrete. Often in the case of continuous distributions it still holds. For example,
1.3. LARGE INSTANCES / BAYESIAN NETWORKS
X
Y
Z
P(x1) = .3
P(x2) = .7
P(y1|x1) = .6
P(y2|x1) = .4
P(z1|y1) = .2
P(z2|y1) = .8
P(y1|x2) = 0
P(y2|x2) = 1
P(z1|y2) = .5
P(z2|x2) = .5
39
Figure 1.6: A DAG containing random variables, along with specified conditional distributions.
it holds for the Gaussian distributions introduced in Section 4.1.3. However,
in general, it does not hold for all continuous conditional distributions. See
[Dawid and Studeny, 1999] for an example in which no joint distribution having
the specified distributions as conditionals even exists.
Example 1.28 Suppose we specify the DAG G shown in Figure 1.6, along with
the conditional distributions shown in that figure. According to Theorem 1.5,
P (x, y, z) = P (z|y)P (y|x)P (x)
satisfies the Markov condition with G.
Note that the proof of Theorem 1.5 does not require that values in the
specified conditional distributions be nonzero. The next example shows what
can happen when we specify some zero values.
Example 1.29 Consider first the DAG and specified conditional distributions
in Figure 1.6. Because we have specified a zero conditional probability, namely
P (y1|x2), there are events in the joint distribution with zero probability. For
example,
P (x2, y1, z1) = P (z1|y1)P (y1|x2)P (x2) = (.2)(0)(.7) = 0.
However, there is no event with zero probability that is a conditioning event in
one of the specified conditional distributions. That is, P (x1), P (x2), P (y1), and
P (y2) are all nonzero. So the specified conditional distributions all exist.
Consider next the DAG and specified conditional distributions in Figure 1.7.
We have
P (x1, y1) = P (x1, y1|w1)P (w1) + P (x1, y1|w2)P (w2)
= P (x1|w1)P (y1|w1)P (w1) + P (x1|w2)P (y1|w2)P (w2)
= (0)(.8)(.1) + (.6)(0)(.9) = 0.
The event x1, y1 is a conditioning event in one of the specified distributions,
namely P (zi|x1, y1), but it has zero probability, which means we can’t condition
40
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
P(w1) = .1
P(w2) = .9
W
P(x1|w1) = 0
P(x2|w1) = 1
P(y1|w1) = .8
P(y2|w1) = .2
X
Y
P(x1|w2) = .6
P(x2|w2) = .4
P(y1|w2) = 0
P(y2|w2) = 1
Z
P(z1|x1,y1) = .3
P(z2|x1,y1) = .7
P(z1|x1,y2) = .4
P(z2|x1,y2) = .6
P(z1|x2,y1) = .1
P(z2|x2,y1) = .9
P(z1|x2,y2) = .5
P(z2|x2,y2) = .5
Figure 1.7: The event x1, y1 has 0 probability.
on it. This poses no problem; it simply means we have specified some meaningless values, namely P (zi|x1, y1). The Markov condition is still satisfied because
P (z|w, x, y) = P (z|x, y) whenever P (x, y) 6= 0 (See the definition of conditional
independence for sets of random variables in Section 1.1.4.).
1.3.3
Bayesian Networks
Let P be a joint probability distribution of the random variables in some set
V, and G = (V, E) be a DAG. We call (G, P ) a Bayesian network if (G, P )
satisfies the Markov condition. Owing to Theorem 1.4, P is the product of its
conditional distributions in G, and this is the way P is always represented in
a Bayesian network. Furthermore, owing to Theorem 1.5, if we specify a DAG
G and any discrete conditional distributions (and many continuous ones), we
obtain a Bayesian network This is the way Bayesian networks are constructed
in practice. Figures 1.5, 1.6, and 1.7 all show Bayesian networks.
Example 1.30 Figure 1.8 shows a Bayesian network containing the probability
distribution discussed in Example 1.23.
Example 1.31 Recall the objects in 1.2 and the resultant joint probability distribution P discussed in Example 1.25. Example 1.27 developed a Bayesian
network (namely the one in Figure 1.5) containing that distribution. Figure 1.9
shows another Bayesian network whose conditional distributions are obtained
1.3. LARGE INSTANCES / BAYESIAN NETWORKS
41
Lung
P(LungCancer = present) = .001
Cancer
P(Test = positive|LungCancer = present) = .6
Test
P(Test = positive|LungCancer = absent) = .02
Figure 1.8: A Bayesian network representing the probability distribution discussed in Example 1.23.
P(v1) = 5/13
P(s1) = 8/13
V
S
C
P(c1|v1,s1) = 2/3
P(c1|v1,s2) = 1/2
P(c1|v2,s1) = 4/5
P(c1|v2,s2) = 2/3
Figure 1.9: A Bayesian network.
42
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
P(h1) = .2
H
P(b1|h1) = .25
P(b1|h2) = .05
B
L
P(l1|h1) = .003
P(l1|h2) = .00005
F
P(f1|b1,l1) = .75
P(f1|b1,l2) = .10
P(f1|b2,l1) = .5
P(f1|b2,l2) = .05
C
P(c1|l1) = .6
P(c1|l2) = .02
Figure 1.10: A Bayesian nework.
from P . Does this Bayesian network contain P ? No it does not. Since P does
not satisfy the Markov condition with the DAG in that figure, there is no reason
to suspect P would be the product of the conditional distributions in that DAG.
It is a simple matter to verify that indeed it is not. So, although the Bayesian
network in Figure 1.9 contains a probability distribution, it is not P .
Example 1.32 Recall the situation discussed at the beginning of this section
where we were concerned with the joint probability distribution of smoking history (H), bronchitis (B), lung cancer (L), fatigue (F ), and chest X-ray (C).
Figure 1.1, which appears again as Figure 1.10, shows a Bayesian network containing those variables in which the conditional distributions were estimated from
actual data.
Does the Bayesian network in the previous example contain the actual relative frequency distribution of the variables? Example 1.31 illustrates that if
we develop a Bayesian network from an arbitrary DAG and the conditionals
of a probability distribution P relative to that DAG, in general the resultant
Bayesian network does not contain P . Notice that, in Figure 1.10 we constructed the DAG using causal edges. For example, there is an edge from H
to L because smoking causes lung cancer. In the next section, we argue that
if we construct a DAG using causal edges we often have a DAG that satisfies
the Markov condition with the relative frequency distribution of the variables.
Given this, owing to Theorem 1.4, the relative frequency distribution of the
variables in Figure 1.10 should satisfy the Markov condition with the DAG in
1.4. CREATING BAYESIAN NETWORKS USING CAUSAL EDGES
43
that figure. However, the situation is different than our urn example (Examples 1.25 and 1.27). Even if the values in the conditional distribution in Figure
1.10 are obtained from relative frequency data, they will only be estimates of
the actual relative frequencies. Therefore, the resultant joint distribution is a
different joint distribution than the joint relative frequency distribution of the
variables. What distribution is it? It is our joint subjective probability distribution P of the variables obtained from our beliefs concerning conditional
independencies among the variables (the structure of the DAG G) and relative
frequency data. Theorem 1.5 tells us that in many cases (G, P ) satisfies the
Markov condition and is therefore a Bayesian network. Note, that if we are
correct about the conditional independencies, we will have convergence to the
actual relative frequency distribution.
1.3.4
A Large Bayesian Network
In this section, we introduced Bayesian networks and we demonstrated their
application using small textbook examples. To illustrate their practical use, we
close by briefly discussing a large-scale Bayesian network used in a system called
NasoNet.
NasoNet [Galán et al, 2002] is a system that performs diagnosis and prognosis of nasopharyngeal cancer, which is cancer concerning the nasal passages.
The Bayesian network used in NasoNet contains 15 nodes associated with tumors confined to the nasopharynx, 23 nodes representing the spread of tumors
to nasopharyngeal surrounding sites, 4 nodes concerning distant metastases, 4
nodes indicating abnormal lymph nodes, 11 nodes expressing nasopharyngeal
hemorrheages or infections, and 50 nodes representing symptoms or syndromes
(combinations of symptoms). Figure 1.11 show a portion of the Bayesian network. The feature shown in each node either has value present or absent.
NasoNet models the evolution of nasopharyngeal cancer in such a way that
each arc represents a causal relation between the parent and the child. For
example, in Figure 1.11 the presence of infection in the nasopharynx may cause
rhinorrhea (excessive mucous secretion from the nose). The next section discusses why constructing a DAG with causal edges should often yield a Bayesian
network.
1.4
Creating Bayesian Networks Using Causal
Edges
Given a set of random variables V, if for every X, Y ∈ V we draw an edge
from X to Y if and only if X is a direct cause of Y relative to V, we call the
resultant DAG a causal DAG. In this section, we illustrate why we feel the
joint probability (relative frequency) distribution of the variables in a causal
DAG often satisfies the Markov condition with that DAG, which means we
can construct a Bayesian network by creating a causal DAG. Furthermore, we
explain what we mean by ‘X is a direct cause of Y relative to V’ (at least for
44
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
Primary
vegetating tumor
on
right lateral wall
Vegetating tumor
occupying
right nasal fossa
Persistent
nasal obstruction
on the right side
Primary infiltrating
tumor on
superior wall
Infection
in the
nasopharynx
Infiltrating tumor
spread to
anterior wall
Rhinorrhea
Infiltrating tumor
spread to
right nasal fossa
Anosmia
Figure 1.11: Part of the Bayesian network in Nasonet.
one definition of causation). Before doing this, we first review the concept of
causation and a method for determining causal influences.
1.4.1
Ascertaining Causal Influences Using Manipulation
Some of what follows is based on a similar discussion in [Cooper, 1999]. One
dictionary definition of a cause is ‘the one, such as a person, an event, or a
condition, that is responsible for an action or a result.’ Although useful, this
simple definition is certainly not the last word on the concept of causation, which
has been wrangled about philosophically for centuries (See e.g. [Eells, 1991],
[Hume, 1748], [Piaget, 1966], [Salmon, 1994], [Spirtes et al, 1993, 2000].). The
definition does, however, shed light on an operational method for identifying
causal relationships. That is, if the action of making variable X take some
value sometimes changes the value taken by variable Y , then we assume X is
1.4. CREATING BAYESIAN NETWORKS USING CAUSAL EDGES
45
responsible for sometimes changing Y ’s value, and we conclude X is a cause of
Y . More formally, we say we manipulate X when we force X to take some
value, and we say X causes Y if there is some manipulation of X that leads to a
change in the probability distribution of Y . We assume that if manipulating X
leads to a change in the probability distribution of Y , then X obtaining a value
by any means whatsoever also leads to a change in the probability distribution of
Y . So we assume causes and their effects are statistically correlated. However,
as we shall discuss soon, variables can be correlated without one causing the
other. A manipulation consists of a randomized controlled experiment (RCE)
using some specific population of entities (e.g. individuals with chest pain) in
some specific context (E.g., they currently receive no chest pain medication and
they live in a particular geographical area.). The causal relationship discovered
is then relative to this population and this context.
Let’s discuss how the manipulation proceeds. We first identify the population of entities we wish to consider. Our random variables are features of
these entities. Next we ascertain the causal relationship we wish to investigate.
Suppose we are trying to determine if variable X is a cause of variable Y . We
then sample a number of entities from the population (See Section 4.2.1 for a
discussion of sampling.). For every entity selected, we manipulate the value of
X so that each of its possible values is given to the same number of entities (If X
is continuous, we choose the values of X according to a uniform distribution.).
After the value of X is set for a given entity, we measure the value of Y for
that entity. The more the resultant data shows a dependency between X and
Y the more the data supports that X causally influences Y . The manipulation
of X can be represented by a variable M that is external to the system being
studied. There is one value mi of M for each value xi of X, the probabilities
of all values of M are the same, and when M equals mi, X equals xi. That is,
the relationship between M and X is deterministic. The data supports that X
causally influences Y to the extent that the data indicates P (yi|mj) 6= P (yi|mk)
for j 6= k. Manipulation is actually a special kind of causal relationship that we
assume exists primordially and is within our control so that we can define and
discover other causal relationships.
An Illustration of Manipulation
We demonstrate these ideas with a comprehensive example concerning recent
headline news. The pharmaceutical company Merck had been marketing its drug
finasteride as medication for men for a medical condition. Based on anecdotal
evidence, it seemed that there was a correlation between use of the drug and
regrowth of scalp hair. Let’s assume that Merck determined such a correlation
does exist. Should they conclude finasteride causes hair regrowth and therefore
market it as a cure for baldness? Not necessarily. There are quite a few causal
explanations for the correlation of two variables. We discus these next.
Possible Causal Relationships Let F be a variable representing finasteride
use and G be a variable representing scalp hair growth. The actual values of F
46
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
F
G
F
(a)
G
(b)
H
F
G
F
(c)
F
G
(d)
G
Y
(e)
Figure 1.12: All five causal relationships could account for F and G being
correlated.
and G are unimportant to the present discussion. We could use either continuous or discrete values. If F caused G, then indeed they would be statistically
correlated, but this would also be the case if G caused F , or if they had some
hidden common cause H. If we again represent a causal influence by a directed
edge, Figure 1.12 shows these three possibilities plus two more. Figure 1.12 (a)
shows the conjecture that F causes G, which we already suspect might be the
case. However, it could be that G causes F (Figure 1.12 (b)). You may argue
that, based on domain knowledge, this does not seem reasonable. However, in
general we do not have domain knowledge when doing a statistical analysis. So
from the correlation alone, the causal relationships in Figure 1.12 (a) and (b)
are equally reasonable. Even in this domain, G causing F seems possible. A
man may have used some other hair regrowth product such as minoxidil, which
caused him to regrow hair, became excited about the regrowth, and decided to
try other products such as finasteride which he heard might cause regrowth. As
a third possibility, it could be both that finasteride causes hair regrowth and
hair regrowth causes use of finasteride, meaning we could have a causal loop or
1.4. CREATING BAYESIAN NETWORKS USING CAUSAL EDGES
47
feedback. Therefore, Figure 1.12 (c) is also a possibility. For example, finasteride may cause regrowth, and excitement about regrowth may cause use of
finasteride. A fourth possibility, shown in Figure 1.12 (d), is that F and G have
some hidden common cause H which accounts for their statistical correlation.
For example, a man concerned about hair loss might try both finasteride and
minoxidil in his effort to regrow hair. The minoxidil may cause hair regrowth,
while the finasteride does not. In this case the man’s concern is a cause of
finasteride use and hair regrowth (indirectly through minoxidil use), while the
latter two are not causally related. A fifth possibility is that we are observing a
population in which all individuals have some (possibly hidden) effect of both
F and G. For example, suppose finasteride and apprehension about lack of
hair regrowth are both causes of hypertension2 , and we happen to be observing
individuals who have hypertension Y . We say a node is instantiated when we
know its value for the entity currently being modeled. So we are saying Y is
instantiated to the same value for all entities in the population we are observing.
This situation is depicted in Figure 1.12 (e), where the cross through Y means
the variable is instantiated. Ordinarily, the instantiation of a common effect
creates a dependency between its causes because each cause explains away the
occurrence of the effect, thereby making the other cause less likely. Psychologists call this discounting. So, if this were the case, discounting would explain
the correlation between F and G. This type of dependency is called selection
bias. A final possibility (not depicted in Figure 1.12) is that F and G are not
causally related at all. The most notable example of this situation is when our
entities are points in time, and our random variables are values of properties
at these different points in time. Such random variables are often correlated
without having any apparent causal connection. For example, if our population
consists of points in time, J is the Dow Jones Average at a given time, and L
is Professor Neapolitan’s hairline at a given time, then J and L are correlated.
Yet they do not seem to be causally connected. Some argue there are hidden
common causes beyond our ability to measure. We will not discuss this issue
further here. We only wish to note the difficulty with such correlations. In light
of all of the above, we see then that we cannot deduce the causal relationship
between two variables from the mere fact that they are statistically correlated.
It may not be obvious why two variables with a common cause would be
correlated. Consider the present example. Suppose H is a common cause of F
and G and neither F nor G caused the other. Then H and F are correlated
because H causes F , H and G are correlated because H causes G, which implies F and G are correlated transitively through H. Here is a more detailed
explanation. For the sake of example, suppose h1 is a value of H that has a
causal influence on F taking value f 1 and on G taking value g1. Then if F
had value f 1, each of its causes would become more probable because one of
them should be responsible. So P (h1|f 1) > P (f 1). Now since the probability
of h1 has gone up, the probability of g1 would also go up because h1 causes g1.
2 There is no evidence that either finasteride or apprenhension about lack of hair regrowth
cause hypertension. This is only for the sake of illustration.
48
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
M
F
P(m1) = .5
P(m2) = .5
P(f1|m1) = 1
P(f2|m1) = 0
G
P(f1|m2) = 0
P(f2|m2) = 1
Figure 1.13: A manipulation investigating whether F causes G.
Therefore, P (g1|f 1) > P (f 1), which means F and G are correlated.
Merck’s Manipulation Study Since Merck could not conclude finasteride
causes hair regrowth from their mere correlation alone, they did a manipulation
study to test this conjecture. The study was done on 1,879 men aged 18 to 41
with mild to moderate hair loss of the vertex and anterior mid-scalp areas. Half
of the men were given 1 mg. of finasteride, while the other half were given 1
mg. of placebo. Let’s define variables for the study, including the manipulation
variable M :
Variable
F
G
M
Value
f1
f2
g1
e2
m1
m2
When the Variable Takes this Value
Subject takes 1 mg. of finasteride.
Subject takes 1 mg. of placebo.
Subject has significant hair regrowth.
Subject does not have significant hair regrowth.
Subject is chosen to take 1mg of finasteride.
Subject is chosen to take 1mg of placebo.
Figure 1.13 shows the conjecture that F causes G and the RCE used to test
this conjecture. There is an oval around the system being modeled to indicate
the manipulation comes from outside that system. The edges in that figure
represent causal influences. The RCE supports the conjecture that F causes G
to the extent that the data support P (g1|m1) 6= P (g1|m2). Merck decided that
‘significant hair regrowth’ would be judged according to the opinion of independent dermatologists. A panel of independent dermatologists evaluated photos
of the men after 24 months of treatment. The panel judged that significant
hair regrowth was demonstrated in 66 percent of men treated with finasteride
compared to 7 percent of men treated with placebo. Basing our probability on
1.4. CREATING BAYESIAN NETWORKS USING CAUSAL EDGES
F
D
49
G
Figure 1.14: A causal DAG depicting that F causes D and D causes G.
these results, we have P (g1|m1) ≈ .67 and P (g1|m2) ≈ .07. In a more analytical
analysis, only 17 percent of men treated with finasteride demonstrated hair loss
(defined as any decrease in hair count from baseline). In contrast, 72 percent of
the placebo group lost hair, as measured by hair count. Merck concluded that
finasteride does indeed cause hair regrowth, and on Dec. 22, 1997 announced
that the U.S. Food and Drug Administration granted marketing clearance to
Propecia(TM) (finasteride 1 mg.) for treatment of male pattern hair loss (androgenetic alopecia), for use in men only. See [McClennan and Markham, 1999]
for more on this.
Causal Mediaries The action of finasteride is well-known. That is, manipulation experiments have shown it significantly inhibits the conversion of testosterone to dihydro-testosterone (DHT) (See e.g. [Cunningham et al, 1995].). So
without performing the study just discussed, Merck could assume finasteride
(F ) has a causal effect on DHT level (D). DHT is believed to be the androgen responsible for hair loss. Suppose we know for certain that a balding man,
whose DHT level was set to zero, would regrow hair. We could then also conclude DHT level (D) has a causal effect on hair growth (G). These two causal
relationships are depicted in Figure 1.14. Could Merck have used these causal
relations to conclude for certain that finasteride would cause hair regrowth and
avoid the expense of their study? No, they could not. Perhaps, a certain minimal level of DHT is necessary for hair loss, more than that minimal level has
no further effect on hair loss, and finasteride is not capable of lowering DHT
level below that level. That is, it may be that finasteride has a causal effect on
DHT level, DHT level has a causal effect on hair growth, and yet finasteride has
no effect on hair growth. If we identify that F causes D and D causes G, and
F and G are probabilistically independent, we say the probability distribution
of the variables is not faithful to the DAG representing their identified causal
relationships. In general, we say (G, P ) satisfies the faithfulness condition if
(G, P ) satisfies the Markov condition and the only conditional independencies in
P are those entailed by the Markov condition. So, if F and G are independent,
the probability distribution does not satisfy the faithfulness condition with the
DAG in Figure 1.14 because this independence is not entailed by the Markov
condition. Faithfulness, along with its role in causal DAGs, is discussed in detail
in Chapter 2.
Notice that if the variable D was not in the DAG in Figure 1.14, and if the
probability distribution did satisfy the faithfulness condition (which we believe
based on Merck’s study), there would be an edge from F directly into G instead
50
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
of the directed path through D. In general, our edges always represent only the
relationships among the identified variables. It seems we can usually conceive
of intermediate, unidentified variables along each edge. Consider the following
example taken from [Spirtes et al, 1993, 2000] [p. 42].
If C is the event of striking a match, and A is the event of the match
catching on fire, and no other events are considered, then C is a
direct cause of A. If, however, we added B; the sulfur on the match
tip achieved sufficient heat to combine with the oxygen, then we
could no longer say that C directly caused A, but rather C directly
caused B and B directly caused A. Accordingly, we say that B is a
causal mediary between C and A if C causes B and B causes A.
Note that, in this intuitive explanation, a variable name is used to stand also
for a value of the variable. For example, A is a variable whose value is on-fire
or not-on-fire, and A is also used to represent that the match is on fire. Clearly,
we can add more causal mediaries. For example, we could add the variable D
representing whether the match tip is abraded by a rough surface. C would then
cause D, which would cause B, etc. We could go much further and describe
the chemical reaction that occurs when sulfur combines with oxygen. Indeed, it
seems we can conceive of a continuum of events in any causal description of a
process. We see then that the set of observable variables is observer dependent.
Apparently, an individual, given a myriad of sensory input, selectively records
discernible events and develops cause-effect relationships among them. Therefore, rather than assuming there is a set of causally related variables out there, it
seems more appropriate to only assume that, in a given context or application,
we identify certain variables and develop a set of causal relationships among
them.
Bad Manipulation
Before discussing causation and the Markov condition, we note some cautionary
procedures of which one must be aware when performing a RCE. First, we
must be careful that we do not inadvertently disturb the system other than the
disturbance done by the manipulation variable M itself. That is, we must be
careful we do not accidentally have any other causal edges into the system being
modeled. The following is an example of this kind of bad manipulation (due to
Greg Cooper [private correspondence]):
Example 1.33 Suppose we want to determine the relative effectiveness of home
treatment and hospital treatment for low-risk pneumonia patients. Consider
those patients of Dr. Welby who are randomized to home treatment, but whom
Dr. Welby normally would have admitted to the hospital. Dr. Welby may give
more instructions to such home-bound patients than he would give to the typical
home-bound patient. These instructions might influence patient outcomes. If
those instructions are not measured, then the RCE may give biased estimates of
the effect of treatment location (home or hospital) on patient outcome. Note, we
1.4. CREATING BAYESIAN NETWORKS USING CAUSAL EDGES
51
are interested in estimating the effect of treatment location on patient outcomes,
everything else being equal. The RCE is actually telling us the effect of treatment
allocation on patient outcomes, which is not of interest here (although it could be
of interest for other reasons). The manipulation of treatment allocation is a bad
manipulation of treatment location because it not only results in a manipulation
M of treatment location, but it also has a causal effect on physicians’ other
actions such as advice given. This is an example of what some call a ‘fat hand’
manipulation, in the sense that one would like to manipulate just one variable,
but one’s hand is so fat that it ends up affecting other variables as well.
Let’s show with a DAG how this RCE inadvertently disturbs the system being modeled other than the disturbance done by M itself. If we let L represent
treatment location, A represent treatment allocation, and M represent the manipulation of treatment location, we have these values:
Variable
L
A
M
Value
l1
l2
a1
a2
m1
m2
When the Variable Takes this Value
Subject is at home
Subject is in hospital
Subject is allocated to be at home
Subject is allocated to be in hospital
Subject is chosen to stay home
Subject is chosen to stay in hospital
Other variables in the system include E representing the doctor’s evaluation of
the patient, T representing the doctor’s treatments and other advice, and O representing patient outcome. Since these variables can have more than two values
and their actual values are not important to the current discussion, we did not
show their values in the table above. Figure 1.15 shows the relationships among
the five variables. Note that A not only results in the desired manipulation, but
there is another edge from A into the system being modeled, namely the edge
into T . This edge is our inadvertent disturbance.
In many studies (whether experimental or observational) it often is difficult,
if not impossible, to blind clinicians (and often patients) to the actions the clinicians have been randomized to take. Thus, a fat hand manipulation is a real
possibility. Drug studies often are an important exception; however, there are
many clinician actions we would like to study besides drug selection.
Besides fat hand manipulation, another kind of bad manipulation would be
if we could not get complete control in setting the value of the variable we
wish to manipulate. This manipulation is bad with respect to what we want to
accomplish with the manipulation.
1.4.2
Causation and the Markov Condition
Recall from the beginning of Section 1.4 we stated the following: Given a set of
variables V, if for every X, Y ∈ V we draw an edge from X to Y if and only if X
is a direct cause of Y relative to V, we call the resultant DAG a causal DAG.
Given the manipulation definition of causation offered earlier, by ‘X being a
52
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
A
T
E
O
M
L
Figure 1.15: The action A has a causal arc into the system other than through
M.
F
D
G
Figure 1.16: The causal relationships if F had a causal influence on G other
than through D.
direct cause of Y relative to V’ we mean that a manipulation of X changes the
probability distribution of Y , and that there is no subset W ⊆ V − {X, Y } such
that if we instantiate the variables in W a manipulation of X no longer changes
the probability distribution of Y . When constructing a causal DAG containing
a set of variables V, we call V ‘our set of observed variables.’ Recall further from
the beginning of Section 1.4 we said we would illustrate why we feel the joint
probability (relative frequency) distribution of the variables in a causal DAG
often satisfies the Markov condition with that DAG. We do that first; then we
state the causal Markov Assumption.
Why Causal DAGs Often Satisfy the Markov Condition
Consider first the situation concerning finasteride, DHT, and hair regrowth discussed in Section 1.4.1. In this case, our set of observed variables V is {F, D, G}.
We learned that finasteride level has a causal influence on DHT level. So we
placed an edge from F to D in Figure 1.14. We learned that DHT level has a
causal influence on hair regrowth. So we placed an edge from D to G in Figure
1.14. We suspected that the causal effect finasteride has on hair regrowth is
only through the lowering of DHT levels. So we did not place an edge from
F to G in Figure 1.14. If there was another causal path from F to G (i.e. if
1.4. CREATING BAYESIAN NETWORKS USING CAUSAL EDGES
53
H
X
Y
Z
Figure 1.17: X and Y are not independent if they have a hidden common cause
H.
F affected G by some means other than by decreasing DHT levels), we would
also place an edge from F to G as shown in Figure 1.16. Assuming the only
causal connection between F and G is as indicated in Fig 1.14, we would feel
that F and G are conditionally independent given D because, once we knew
the value of D, we would have a probability distribution of G based on this
known value, and, since the value of F cannot change the known value of D and
there is no other connection between F and G, it cannot change the probability
distribution of G. Manipulation experiments have substantiated this intuition.
That is, there have been experiments in which it was established that X causes
Y , Y causes Z, X and Z are not probabilistically independent, and X and Z
are conditionally independent given Y . See [Lugg et al, 1995] for an example.
In general, when all causal paths from X to Y contain at least one variable in
our set of observed variables V, X and Y do not have a common cause, there
are no causal paths from Y back to X, and we do not have selection bias, then
we feel X and Y are independent if we condition on a set of variables including
at least one variable in each of the causal paths from X to Y . Since the set of
all parents of Y is such a set, we feel the Markov condition is satisfied relative
to X and Y .
We say X and Y have a common cause if there is some variable that has
causal paths into both X and Y . If X and Y have a common cause C, there is
often a dependency between them through this common cause (But this is not
necessarily the case. See Exercise 2.34.). However, if we condition on Y ’s parent
in the path from C to Y , we feel we break this dependency for the same reasons
discussed above. So, as long as all common causes are in our set of observed
variables V, we can still break the dependency between X and Y (assuming as
above there are no causal paths from Y to X) by conditioning on the set of
parents of Y , which means the Markov condition is still satisfied relative to X
and Y . A problem arises when at least one common cause is not in our set of
54
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
observed variables V. Such a common cause is called a hidden variable. If two
variables had a hidden common cause, then there would often be a dependency
between them, which the Markov condition would identify as an independency.
For example, consider the DAG in Figure 1.17. If we only identified the variables
X, Y , and Z, and the causal relationships that X and Y each caused Z, we
would draw edges from each of X and Y to Z. The Markov condition would
entail X and Y are independent. But if X and Y had a hidden common cause
H, they would not ordinarily be independent. So, for us to assume the Markov
condition is satisfied, either no two variables in the set of observed variables V
can have a hidden common cause, or, if they do, it must have the same unknown
value for every unit in the population under consideration. When this is the
case, we say the set is causally sufficient.
Another violation of the Markov condition, similar to the failure to include a
hidden common cause, is when there is selection bias present. Recall that, in the
beginning of Section 1.4.1, we noted that if finasteride use (F ) and apprehension
about lack of hair regrowth (G) are both causes of hypertension (Y ), and we
happen to be observing individuals hospitalized for treatment of hypertension,
we would observe a probabilistic dependence between F and G due to selection
bias. This situation is depicted in Figure 1.12 (e). Note that in this situation our
set of observed variables V is {F, G}. That is, Y is not observed. So if neither
F nor G caused each other and they did not have a hidden common cause, a
causal DAG containing only the two variables (i.e. one with no edges) would
still not satisfy the Markov condition with the observed probability distribution,
because the Markov condition says F and G are independent when indeed they
are not for this population.
Finally, we must also make certain that if X has a causal influence on Y , then
Y does not have a causal influence X. In this way we guarantee that the identified causal edges will indeed yield a DAG. Causal feedback loops (e.g. the situation identified in Figure 1.12 (c)) are discussed in [Richardson and Spirtes, 1999].
Before closing, we note that if we mistakenly draw an edge from X to Y
in a case where X’s causal influence on Y is only through other variables in
the model, we have not done anything to thwart the Markov condition being
satisfied. For example, consider again the variables in Figure 1.14. If F ’s only
influence on G was through D, we would not thwart the Markov condition by
drawing an edge from F to G. That is, this does not result in the structure of
the DAG entailing any conditional independencies that are not there. Indeed,
the opposite has happened. That is, the DAG fails to entail a conditional independency (namely I({F }, {G}|{D})) that is there. This is a violation of the
faithfulness condition (discussed in Chapter 2), not the Markov condition. In
general, we would not want to do this because it makes the DAG less informative
and unnecessarily increases the size of the instance (which is important because,
as we shall see in Section 3.6, the problem of doing inference in Bayesian networks is #P -complete). However, a few mistakes of this sort are not that serious
as we can still expect the Markov condition to be satisfied.
1.4. CREATING BAYESIAN NETWORKS USING CAUSAL EDGES
55
The Causal Markov Assumption
We’ve offered a definition of causation based on manipulation, and we’ve argued
that, given this definition of causation, a causal DAG often satisfies the Markov
condition with the probability distribution of the variables, which means we
can construct a Bayesian network by creating a causal DAG. In general, given
any definitions of ‘causation’ and ‘direct causal influence,’ if we create a causal
DAG G = (V, E) and assume the probability distribution of the variables in
V satisfies the Markov condition with G, we say we are making the causal
Markov assumption.
As discussed above, if the following three conditions are satisfied the causal
Markov assumption is ordinarily warranted: 1) there must be no hidden common
causes; 2) selection bias must not be present; and 3) there must be no causal
feedback loops. In general, when constructing a Bayesian network using identified causal influences, one must take care that the causal Markov assumptions
holds.
Often we identify causes using methods other than manipulation. For example, most of us believe smoking causes lung cancer. Yet we have not manipulated
individuals by making them smoke. We believe in this causal influence because
smoking and lung cancer are correlated, the smoking precedes the cancer in time
(a common assumption is that an effect cannot precede a cause), and there are
biochemical changes associated with smoking. All of this could possibly be explained by a hidden common cause (Perhaps a genetic defect causes both.), but
domain experts essentially rule out this possibility. When we identify causes by
any means whatsoever, ordinarily we feel they are ones that could be identified
by manipulation if we were to perform a RCE, and we make the causal Markov
assumption as long as we are confident exceptions such as conditions (1), (2)
and (3) in the preceding paragraph are not present.
An example of constructing a causal DAG follows.
Example 1.34 Suppose we have identified the following causal influences by
some means: A history of smoking (H) has a causal effect both on bronchitis
(B) and on lung cancer (L). Furthermore, each of these variables can cause
fatigue (F ). Lung Cancer (L) can cause a positive chest X-ray (C). Then
the DAG in Figure 1.10 represents our identified causal relationships among
these variables. If we believe 1) these are the only causal influences among the
variables; 2) there are no hidden common causes; and 3) selection bias is not
present, it seems reasonable to make the causal Markov assumption. Then if
the conditional distributions specified in Figure 1.10 are our estimates of the
conditional relative frequencies, that DAG along with those specified conditional
distributions constitute a Bayesian network which represents our beliefs.
Before closing we mention an objection to the causal Markov condition. That
is, unless we abandon the ‘locality principle’ the condition seems to be violated
in some quantum mechanical experiments. See [Spirtes et al, 1993, 2000] for a
discussion of this matter.
56
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
C
V
S
C
(a)
V
S
(b)
Figure 1.18: C and S are not independent in (a), but the instantiation of V in
(b) renders them independent.
The Markov Condition Without Causation
Using causal edges is just one way to develop a DAG and a probability distribution that satisfy the Markov condition. In Example 1.25 we showed the
joint distribution of V (value), S (shape), and C (color) satisfied the Markov
condition with the DAG in Figure 1.3 (a), but we would not say that the color
of an object has a causal influence on its shape. The Markov condition is simply
a property of the probabilistic relationships among the variables. Furthermore,
if the DAG in Figure 1.3 (a) did capture the causal relationships among some
causally sufficient set of variables and there was no selection bias present, the
Markov condition would be satisfied not only with that DAG but also with the
DAGS in Figures 1.3 (b) and (c). Yet we certainly would not say the edges in
these latter two DAGs represent causal influence.
Some Final Examples
To solidify the notion that the Markov condition is often satisfied by a causal
DAG, we close with three simple examples. We present these examples using
an intuitive approach, which shows how humans reason qualitatively with the
dependencies and conditional independencies among variables. In accordance
with this approach, we again use the name of a variable to stand also for a value.
For example, in modeling whether an individual has a cold, we use a variable C
whose value is present or absent, and we also use C to represent that a cold is
present.
Example 1.35 If Alice’s husband Ralph was planning a surprise birthday party
for Alice with a caterer (C), this may cause him to visit the caterer’s store (V ).
The act of visiting that store could cause him to be seen (S) visiting the store.
So the causal relationships among the variables are the ones shown in Figure
1.18 (a). There is no direct path from C to S because planning the party with
the caterer could only cause him to be seen visiting the store if it caused him
to actually visit the store. If Alice’s friend Trixie reported to her that she had
seen Ralph visiting the caterer’s store today, Alice would conclude that he may
be planning a surprise birthday party because she would feel there is a good
chance Trixie really did see Ralph visiting the store, and, if this actually was
the case, there is a chance he may be planning a surprise birthday party. So C
1.4. CREATING BAYESIAN NETWORKS USING CAUSAL EDGES
C
R
C
S
R
(a)
S
(b)
C
H
C
H
R
S
R
S
(c)
57
(d)
Figure 1.19: If C is the only common cause of R and S (a), we need to instantiate only C (b) to render them independent. If they have exactly two common
causes, C and H (c), we need to instantiate both C and H (d) to render them
independent.
and S are not independent. If, however, Alice had already witnessed this same
act of Ralph visiting the caterer’s store, she would already suspect Ralph may be
planning a surprise birthday party. Trixie’s testimony would not affect here belief
concerning Ralph’s visiting the store and therefore would have no affect on her
belief concerning his planning a party. So C and S are conditionally independent
given V , as the Markov condition entails for the DAG in Figure 1.18 (a). The
instantiation of V , which renders C and S independent, is depicted in Figure
1.18 (b) by placing a cross through V .
Example 1.36 A cold (C) can cause both sneezing (S) and a runny nose (R).
Assume neither of these manifestations causes the other and, for the moment,
also assume there are no hidden common causes (That is, this set of variables
is causally sufficient.). The causal relationships among the variables are then
the ones depicted in Figure 1.19 (a). Suppose now that Professor Patel walks
into the classroom with a runny nose. You would fear she has a cold, and,
if so, the cold may make her sneeze. So you back off from her to avoid the
possible sneeze. We see then that S and R are not independent. Suppose next
that Professor Patel calls school in the morning to announce she has a cold
which will make her late for class. When she finally does arrive, you back off
immediately because you feel the cold may make her sneeze. If you see that
58
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
B
F
B
F
A
A
(a)
(b)
Figure 1.20: B and F are independent in (a), but the instantiation of A in (b)
renders them dependent.
her nose is running, this has no affect on your belief concerning her sneezing
because the runny nose no longer makes the cold more probable (You know she
has a cold.). So S and R are conditionally independent given C, as the Markov
condition entails for the DAG in Figure 1.19 (a). The instantiation of C is
depicted in Figure 1.19 (b).
There actually is at least one other common cause of sneezing and a runny
nose, namely hay fever (H). Suppose this is the only common cause missing
from Figure 1.19 (a). The causal relationships among the variables would then
be as depicted in Figure 1.19 (c). Given this, conditioning on C is not sufficient
to render R and S independent, because R could still make S more probable by
making H more probable. So we must condition on both C and H to render R
and S independent. The instantiation of C and H is depicted in Figure 1.19
(d).
Example 1.37 Antonio has observed that his burglar alarm (A) has sometimes
gone off when a freight truck (F ) was making a delivery to the Home Depot in
back of his house. So he feels a freight truck can trigger the alarm. However,
he also believes a burglar (B) can trigger the alarm. He does not feel that
the appearance of a burglar might cause a freight truck to make a delivery or
vice versa. Therefore, he feels that the causal relationships among the variables
are the ones depicted in Figure 1.20 (a). Suppose Antonio sees a freight truck
making a delivery in back of his house. This does not make him feel a burglar
is more probable. So F and B are independent, as the Markov condition entails
for the DAG in Figure 1.20 (a). Suppose next that Antonio is awakened at
night by the sounding of his burglar alarm. This increases his belief that a
burglar is present, and he begins fearing this is indeed the case. However, as he
proceeds to investigate this possibility, he notices that a freight truck is making
a delivery in back of his house. He reasons that this truck explains away the
alarm, and therefore he believes a burglar probably is not present. So he relaxes
a bit. Given the alarm has sounded, learning that a freight truck is present
decreases the probability of a burglar. So the instantiation of A, as depicted in
EXERCISES
59
Figure 1.20 (b), renders F and B conditionally dependent. As noted previously,
the instantiation of a common effect creates a dependence between its causes
because each explains away the occurrence of the effect, thereby making the other
cause less likely.
Note that the Markov condition does not entail that F and B are conditionally dependent given A. Indeed, a probability distribution can satisfy the Markov
condition for a DAG (See Exercise 2.18) without this conditional dependence occurring. However, if this conditional dependence does not occur, the distribution
does not satisfy the faithfulness condition with the DAG. Faithfulness is defined
earlier in this section and is discussed in Chapter 2.
EXERCISES
Section 1.1
Exercise 1.1 Kerrich [1946] performed experiments such as tossing a coin
many times, and he found that the relative frequency did appear to approach a
limit. That is, for example, he found that after 100 tosses the relative frequency
may have been .51, after 1000 it may have been .508, after 10, 000 tosses it may
have been .5003, and after 100, 000 tosses, it may have been .50008. The pattern
is that the 5 in the first place to the right of the decimal point remains in all
relative frequencies after the first 100 tosses, the 0 in the second place remains
in all relative frequencies after the first 1000 tosses, etc. Toss a thumbtack at
least 1000 times and see if you obtain similar results.
Exercise 1.2 Pick some upcoming event (It could be a sporting event or it
could even be the event that you get an ‘A’ in this course.) and determine your
probability of the event using Lindley’s [1985] method of comparing the uncertain
event to a draw of a ball from an urn (See Example 1.3.).
Exercise 1.3 Prove Theorem 1.1.
Exercise 1.4 Example 1.6 showed that, in the draw of the top card from a deck,
the event Queen is independent of the event Spade. That is, it showed P (Queen|
Spade) = P (Queen).
1. Show directly that the event Spade is independent of the event Queen. That
is, show P (Spade|Queen) = P (Spade). Show also that P (Queen∩Spade) =
P (Queen)P (Spade).
2. Show, in general, that if P (E) 6= 0 and P (F) 6= 0, then P (E|F) = P (E) if
and only if P (F|E) = P (F) and each of these holds if and only if P (E∩F) =
P (E)P (F).
60
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
Exercise 1.5 The complement of a set E consists of all the elements in Ω that
are not in E and is denoted by E.
1. Show that E is independent of F if and only if E is independent of F, which
is true if and only if E is independent of F.
2. Example 1.8 showed that, for the objects in Figure 1.2, One and Square are
conditionally independent given Black and given White. Let Two be the set
of all objects containing a ‘2’ and Round be the set of all round objects.
Use the result just obtained to conclude Two and Square, One and Round,
and Two and Round are each conditionally independent given either Black
or White.
Exercise 1.6 Example 1.7 showed that, in the draw of the top card from a
deck, the event E = {kh, ks, qh} and the event F = {kh, kc, qh} are conditionally
independent given the event G = {kh, ks, kc, kd}. Determine whether E and F
are conditionally independent given G.
Exercise 1.7 Prove the rule of total probability, which says if we have n mutually exclusive and exhaustive events E1 , E2 , . . . En , then for any other event
F,
n
X
P (F ∩ Ei ).
P (F) =
i=1
Exercise 1.8 Let Ω be the set of all objects in Figure 1.2, and assign each
object a probability of 1/13. Let One be the set of all objects containing a 1,
and Square be the set of all square objects. Compute P (One|Square) directly and
using Bayes’ Theorem.
Exercise 1.9 Let a joint probability distribution be given. Using the law of
total probability, show that the probability distribution of any one of the random
variables is obtained by summing over all values of the other variables.
Exercise 1.10 Use the results in Exercise 1.5 (1) to conclude that it was only
necessary in Example 1.18 to show that P (r, t) = P (r, t|s1) for all values of r
and t.
Exercise 1.11 Suppose we have two random variables X and Y with spaces
{x1, x2} and {y1, y2} respectively.
1. Use the results in Exercise 1.5 (1) to conclude that we need only show
P (y1|x1) = P (y1) to conclude IP (X, Y ).
2. Develop an example showing that if X and Y both have spaces containing
more than two values, then we need check whether P (y|x) = P (y) for all
values of x and y to conclude IP (X, Y ).
Exercise 1.12 Consider the probability space and random variables given in
Example 1.17.
EXERCISES
61
1. Determine the joint distributions of S and W , of W and H, and the
remaining values in the joint distribution of S, H, and W .
2. Show that the joint distribution of S and H can be obtained by summing
the joint distribution of S, H, and W over all values of W .
3. Are H and W independent? Are H and W conditionally independent
given S? If this small sample is indicative of the probabilistic relationships
among the variables in some population, what causal relationships might
account for this dependency and conditional independency?
Exercise 1.13 The chain rule states that given n random variables X1 , X2 , . . .
Xn , defined on the same sample space Ω,
P (x1 , x2 , . . .xn ) = P (xn |xn−1 , xn−2 , . . .x1 ) · · · P (x2 |x1 )P (x1 )
whenever P (x1 , x2 , . . .xn ) 6= 0. Prove this rule.
Section 1.2
Exercise 1.14 Suppose we are developing a system for diagnosing viral infections, and one of our random variables is F ever. If we specify the possible values
yes and no, is the clarity test passed? If not, further distinguish the values so
it is passed.
Exercise 1.15 Prove Theorem 1.3.
Exercise 1.16 Let V = {X, Y, Z}, let X, Y , and Z have spaces {x1, x2},
{y1, y2}, and {z1, z2} respectively, and specify the following values:
P (x1) = .2
P (x2) = .8
P (y1|x1) = .3
P (y2|x1) = .7
P (z1|x1) = .1
P (z2|x1) = .9
P (y1|x2) = .4
P (y2|x2) = .6
P (z1|x2) = .5
P (z2|x2) = .5.
Define a joint probability distribution P of X, Y , and Z as the product of these
values.
1. Show that the values in this joint distribution sum to 1, and therefore this
is a way of specifying a joint probability distribution according to Definition
1.8.
2. Show further that IP (Z, Y |X). Note that this conditional independency
follows from Theorem 1.5 in Section 1.3.3.
62
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
Exercise 1.17 A forgetful nurse is supposed to give Mr. Nguyen a pill each
day. The probability that she will forget to give the pill on a given day is .3. If
he receives the pill, the probability he will die is .1. If he does not receive the pill,
the probability he will die is .8. Mr. Nguyen died today. Use Bayes’ theorem to
compute the probability the nurse forgot to give him the pill.
Exercise 1.18 An oil well may be drilled on Professor Neapolitan’s farm in
Texas. Based on what has happened on similar farms, we judge the probability
of oil being present to be .5, the probability of only natural gas being present to
be .2, and the probability of neither being present to be .3. If oil is present, a
geological test will give a positive result with probability .9; if only natural gas
is present, it will give a positive result with probability .3; and if neither are
present, the test will be positive with probability .1. Suppose the test comes back
positive. Use Bayes’ theorem to compute the probability oil is present.
Section 1.3
Exercise 1.19 Consider Figure 1.3.
1. The probability distribution in Example 1.25 satisfies the Markov condition
with the DAGs in Figures 1.3 (b) and (c). Therefore, owing to Theorem
1.4, that probability distribution is equal to the product of its conditional
distributions for each of them. Show this directly.
2. Show that the probability distribution in Example 1.25 is not equal to the
product of its conditional distributions for the DAG in Figure 1.3 (d).
Exercise 1.20 Create an arrangement of objects similar to the one in Figure
1.2, but with a different distribution of values, shapes, and colors, so that, if
random variables V , S, and C are defined as in Example 1.25, then the only
independency or conditional independency among the variables is IP (V, S). Does
this distribution satisfy the Markov condition with any of the DAGs in Figure
1.3? If so, which one(s)?
Exercise 1.21 Complete the proof of Theorem 1.5 by showing the specified conditional distributions are the conditional distributions they notationally represent
in the joint distribution.
Exercise 1.22 Consider the objects in Figure 1.2 and the random variables
defined in Example 1.25. Repeatedly sample objects with replacement to obtain
estimates of P (c), P (v|c), and P (s|c). Take the product of these estimates and
compare it to the actual joint probability distribution.
EXERCISES
63
Exercise 1.23 Consider the objects in Figure 1.2 and the joint probability distribution of the random variables defined in Example 1.25. Suppose we compute
its conditional distributions for the DAG in Figure 1.3 (d), and we take their
product. Theorem 1.5 says this product is a joint probability distribution that
constitutes a Bayesian network with that DAG. Is this the actual joint probability
distribution of the variables? If not, what is it?
Section 1.4
Exercise 1.24 Professor Morris investigated gender bias in hiring in the following way. He gave hiring personnel equal numbers of male and female resumes
to review, and then he investigated whether their evaluations were correlated with
gender. When he submitted a paper summarizing his results to a psychology
journal, the reviewers rejected the paper because they said this was an example
of fat hand manipulation. Explain why they might have thought this. Elucidate
your explanation by identifying all relevant variables in the RCE and drawing a
DAG like the one in Figure 1.15.
Exercise 1.25 Consider the following piece of medical knowledge taken from
[Lauritzen and Spiegelhalter, 1988]: Tuberculosis and lung cancer can each cause
shortness of breath (dyspnea) and a positive chest X-ray. Bronchitis is another
cause of dyspnea. A recent visit to Asia can increase the probability of tuberculosis. Smoking can cause both lung cancer and bronchitis. Create a DAG
representing the causal relationships among these variables. Complete the construction of a Bayesian network by determining values for the conditional probability distributions in this DAG either based on your own subjective judgement
or from data.
64
CHAPTER 1. INTRODUCTION TO BAYESIAN NETWORKS
Chapter 2
More DAG/Probability
Relationships
The previous chapter only introduced one relationship between probability distributions and DAGs, namely the Markov condition. However, the Markov condition only entails independencies; it does not entail any dependencies. That is,
when we only know that (G, P ) satisfies the Markov condition, we know the absence of an edge between X any Y entails there is no direct dependency between
X any Y , but the presence of an edge between X and Y does not mean there is a
direct dependency. In general, we would want an edge to mean there is a direct
dependency. In Section 2.3, we discuss another condition, namely the faithfulness condition, which does entail this. The concept of faithfulness is essential to
the methods for learning the structure of Bayesian networks from data, which
are discussed in Chapters 8-11. For some probability distributions P it is not
possible to find a DAG with which P satisfies the faithfulness condition. In Section 2.4 we present the minimality condition, and we shall see that it is always
possible to find a DAG G such that (G, P ) satisfies the minimality condition. In
Section 2.5 we discuss Markov blankets and Markov boundaries, which are sets
of variables that render a given variable conditionally independent of all other
variables. Finally, in Section 2.6 we show how the concepts addressed in this
chapter relate to causal DAGs. Before any of this, in Section 2.1 we show what
conditional independencies are entailed by the Markov condition, and in Section 2.2 we describe Markov equivalence, which groups DAGs into equivalence
classes based on the conditional independencies they entail. Knowledge of the
conditional independencies entailed by the Markov condition is needed to develop a message-passing inference algorithm in Chapter 3, while the concept of
Markov equivalence is necessary to the structure learning algorithms developed
in Chapters 8-11.
65
66
2.1
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
Entailed Conditional Independencies
If (G, P ) satisfies the Markov condition, then each node in G is conditionally
independent of the set of all its nondescendents given its parents. Do these
conditional independencies entail any other conditional independencies? That
is, if (G, P ) satisfies the Markov condition, are there any other conditional independencies which P must satisfy other than the one based on a node’s parents?
The answer is yes. Before explicitly stating these entailed independencies, we
illustrate that one would expect them.
First we make the notion of ‘entailed conditional independency’ explicit:
Definition 2.1 Let G = (V, E) be a DAG, where V is a set of random variables.
We say that, based on the Markov condition, G entails conditional independency IP (A, B|C) for A, B, C ⊆ V if
IP (A, B|C) holds for every P ∈ PG ,
where PG is the set of all probability distributions P such that (G, P ) satisfies
the Markov condition. We also say the Markov condition entails the conditional
independency for G and that the conditional independency is in G.
Note that the independency IP (A, B) is included in the previous definition
because it is the same as IP (A, B|∅). Regardless of whether C is the empty set,
for brevity we often just refer to IP (A, B|C) as an ‘independency’ instead of a
‘conditional independency’.
2.1.1
Examples of Entailed Conditional Independencies
Suppose some distribution P satisfies the Markov condition with the DAG
in Figure 2.1. Then we know IP ({C}, {F, G}|{B}) because B is the parent of C, and F and G are nondescendents of C. Furthermore, we know
IP ({B}, {G}|{F }) because F is the parent of B, and G is a nondescendent
of B. These are the only conditional independencies according to the statement
of the Markov condition. However, can any other conditional independencies be
deduced from them? For example, can we conclude IP ({C}, {G}|{F })? Let’s
first give the variables meaning and the DAG a causal interpretation to see if
we would expect this conditional independency.
Suppose we are investigating how professors obtain citations, and the variables represent the following:
G:
F:
B:
C:
Graduate Program Quality
First Job Quality
Number of Publications
Number of Citations.
Further suppose the DAG in Figure 2.1 represents the causal relationships
among these variable, there are no hidden common causes, and selection bias is
2.1. ENTAILED CONDITIONAL INDEPENDENCIES
67
G
F
B
C
Figure 2.1: I({C}, {G}|{F }) can be deduced from the Markov condition.
not present.1 Then it is reasonable to make the causal Markov assumption, and
we would feel the probability distribution of the variables satisfies the Markov
condition with the DAG. Given all this, if we learned that Professor La Budde
attended a graduate program of high quality (That is, we found out the value
of G for Professor La Budde was ‘high quality’.), we would expect his first job
may well be of high quality, which means there should be a large number of
publications, which in turn implies there should be a large number of citations.
Therefore, we would not expect IP (C, G). If we learned that Professor Pellegrini’s first job was of the high quality (That is, we found out the value of F for
Professor Pellegrini was ‘high quality’.), we would expect his number of publications to be large, and in turn his number of citations to be large. That is, we
would also not expect IP (C, F ). If Professor Pellegrini then told us he attended
a graduate program of high quality, would we expect the number of citations
to be even higher than we previously thought? It seems not. The graduate
program’s high quality implies the number of citations is probably large because it implies the first job is probably of high quality. Once we already know
the first job is of high quality, the information on the graduate program should
be irrelevant to our beliefs concerning the number of citations. Therefore, we
would expect C to not only be conditionally independent of G given its parent
B, but also its grandparent F . Either one seems to block the dependency be1 We make no claim this model accurately represents the causal relationships among the
variables. See [Spirtes et al, 1993, 2000] for a detailed discussion of this problem.
68
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
A
G
F
B
C
Figure 2.2: IP ({C}, {G}|{A, F }) can be deduced from the Markov condition.
tween G and C that exists through the chain [G, F, B, C]. So we would expect
IP ({C}, {G}|{F }).
It is straightforward to show that the Markov condition does indeed entail
IP ({C}, {G}|{F }) for the DAG G in Figure 2.1. We illustrate this for the case
where the variables are discrete. If (G, P ) satisfies the Markov condition,
X
P (c|g, f ) =
P (c|b, g, f )P (b|g, f )
b
=
X
P (c|b, f)P (b|f )
b
= P (c|f ).
The second step is due to the Markov condition.
Suppose next we have an arbitrarily long directed linked list of variables
and P satisfies the Markov condition with that list. In the same way as above,
we can show that, for any variable in the list, the set of variables above it are
conditionally independent of the set of variables below it given that variable.
Suppose now that P does not satisfy the Markov condition with the DAG
in Figure 2.1 because there is a common cause A of G and B. For the sake of
2.1. ENTAILED CONDITIONAL INDEPENDENCIES
69
A
G
F
B
C
Figure 2.3: The Markov condition does not entail I({F }, {A}|{B, G}).
illustration, let’s say A represents the following in the current example:
A:
Ability.
Further suppose there are no other hidden common causes so that we would now
expect P to satisfy the Markov condition with the DAG in Figure 2.2. Would we
still expect IP ({C}, {G}|{F })? It seems not. For example, suppose again that
we initially learn Professor Pellegrini’s first job was of high quality. As before,
we would feel it probable that he has a high number of citations. Suppose again
that we next learn his graduate program was of high quality. Given the current
model, this fact is indicative of his having high ability, which can affect his
publication rate (and thereby his citation rate) directly. So we would not feel
IP ({C}, {G}|{F }) as we did with the previous model. However, if we knew the
state of Professor Pellegrini’s ability, his attendance at a high quality graduate
program could no longer be indicative of his ability, and therefore it would not
affect our belief concerning his citation rate through the chain [G, A, B, C]. That
is, this chain is blocked at A. So we would expect IP ({C}, {G}|{A, F }). Indeed,
it is possible to prove the Markov condition does entail IP ({C}, {G}|{A, F }) for
the DAG in Figure 2.2.
Finally, consider the conditional independency IP ({F }, {A}|{G}). This independency is obtained directly by applying the Markov condition to the DAG
70
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
X
Y
Z
Figure 2.4: There is an uncoupled head-to-head meeting at Z.
in Figure 2.2. So we will not offer an intuitive explanation for it. Rather we
discuss whether we would expect the independency to be maintained if we also
learned the state of B. That is, would we expect IP ({F }, {A}|{B, G})? Suppose we first learn Professor Georgakis has a high publication rate (the value
of B) and attended a high quality graduate program (the value of G). Then
we later learned she also has high ability (the value of A). In this case, her
high ability could explain away her high publication rate, thereby making it less
probable she had a high quality first job (As mentioned in Section 1.4.1, psychologists call this explaining away discounting.) So the chain [A, B, F ] is opened
by instantiating B, and we would not expect IP ({F }, {A}|{B, G}). Indeed, the
Markov condition does not entail IP ({F }, {A}|{B, G}) for the DAG in Figure
2.2. This situation is illustrated in Figure 2.3. Note that the instantiation of
C should also open the chain [A, B, F ]. That is, if we know the citation rate is
high, then it is probable the publication rate is high, and each of the causes of B
can explain away this high probability. Indeed, the Markov condition does not
entail IP ({F }, {A}|{C, G}) either. Note further that we are only saying that the
Markov condition doe not entail IP ({F }, {A}|{B, G}). We are not saying the
Markov condition entails qIP ({F }, {A}|{B, G}). Indeed, the Markov condition
can never entail a dependency; it can only entail an independency. Exercise 2.18
shows an example where this conditional dependency does not occur. That is,
it shows a case where there is no discounting.
2.1.2
d-Separation
We showed in Section 2.1.1 that the Markov condition entails IP ({C}, {G}|{F })
for the DAG in Figure 2.1. This conditional independency is an example of
a DAG property called ‘d-separation’. That is, {C} and {G} are d-separated
by {A, F } in the DAG in Figure 2.1. Next we develop the concept of dseparation, and we show the following: 1) The Markov condition entails that
all d-separations are conditional independencies; and 2) every conditional independencies entailed by the Markov condition is identified by d-separation. That
is, if (G, P ) satisfies the Markov condition, every d-separation in G is a conditional independency in P . Furthermore, every conditional independency, which
is common to all probability distributions satisfying the Markov condition with
2.1. ENTAILED CONDITIONAL INDEPENDENCIES
71
G, is identified by d-separation.
All d-separations are Conditional Independencies
First we need review more graph theory. Suppose we have a DAG G = (V, E),
and a set of nodes {X1 , X2 , . . . ., Xk }, where k ≥ 2, such (Xi−1 , Xi ) ∈ E or
(Xi , Xi−1 ) ∈ E for 2 ≤ i ≤ k. We call the set of edges connecting the k
nodes a chain between X1 and Xk . We denote the chain using both the sequence [X1 , X2 , . . . ., Xk ] and the sequence [Xk , Xk−1 , . . . ., X1 ]. For example,
[G, A, B, C] and [C, B, A, G] represent the same chain between G and C in the
DAG in Figure 2.3. Another chain between G and C is [G, F, B, C]. The nodes
X2 , . . . Xk−1 are called interior nodes on chain [X1 , X2 , . . . Xk ]. The subchain
of chain [X1 , X2 , . . . Xk ] between Xi and Xj is the chain [Xi , Xi+1 , . . . Xj ] where
1 ≤ i < j ≤ k. A cycle is a chain between a node and itself. A simple chain
is a chain containing no subchains which are cycles. We often denote chains
by showing undirected lines between the nodes in the chain. For example, we
would denote the chain [G, A, B, C] as G − A − B − C. If we want to show the
direction of the edges, we use arrows. For example, to show the direction of the
edges, we denote the previous chain as G ← A → B → C. A chain containing
two nodes, such as X − Y , is called a link. A directed link, such as X → Y ,
represents an edge, and we will call it an edge. Given the edge X → Y , we say
the tail of the edge is at X and the head of the edge is Y . We also say the
following:
• A chain X → Z → Y is a head-to—tail meeting, the edges meet headto-tail at Z, and Z is a head-to-tail node on the chain.
• A chain X ← Z → Y is a tail-to—tail meeting, the edges meet tail-totail at Z, and Z is a tail-to-tail node on the chain.
• A chain X → Z ← Y is a head-to—head meeting, the edges meet
head-to-head at Z, and Z is a head-to-head node on the chain.
• A chain X − Z − Y , such that X and Y are not adjacent, is an uncoupled
meeting.
Figure 2.4 shows an uncoupled head-to-head meeting.
We now have the following definition:
Definition 2.2 Let G = (V, E) be a DAG, A ⊆ V, X and Y be distinct nodes
in V − A, and ρ be a chain between X and Y . Then ρ is blocked by A if one of
the following holds:
1. There is a node Z ∈ A on the chain ρ, and the edges incident to Z on ρ
meet head-to-tail at Z.
2. There is a node Z ∈ A on the chain ρ, and the edges incident to Z on ρ
meet tail-to-tail at Z.
72
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
X
W
Y
Z
R
S
T
Figure 2.5: A DAG used to illustrate chain blocking.
3. There is a node Z, such that Z and all of Z’s descendents are not in A,
on the chain ρ, and the edges incident to Z on ρ meet head-to-head at Z.
We say the chain is blocked at any node in A where one of the above meetings
takes place. There may be more than one such node. The chain is called active
given A if it is not blocked by A.
Example 2.1 Consider the DAG in Figure 2.5.
1. The chain [Y, X, Z, S] is blocked by {X} because the edges on the chain
incident to X meet tail-to-tail at X. That chain is also blocked by {Z}
because the edges on the chain incident to Z meet head-to-tail at Z.
2. The chain [W, Y, R, Z, S] is blocked by ∅ because R ∈
/ ∅, T ∈
/ ∅, and the
edges on the chain incident to R meet head-to-head at R.
3. The chain [W, Y, R, S] is blocked by {R} because the edges on the chain
incident to R meet head-to-tail at R.
4. The chain [W, Y, R, Z, S] is not blocked by {R} because the edges on the
chain incident to R meet head-to-head at R. Furthermore, this chain is
not blocked by {T } because T is a descendent of R.
We can now define d-separation.
Definition 2.3 Let G = (V, E) be a DAG, A ⊆ V, and X and Y be distinct
nodes in V − A. We say X and Y are d-separated by A in G if every chain
between X and Y is blocked by A.
2.1. ENTAILED CONDITIONAL INDEPENDENCIES
73
It is not hard to see that every chain between X and Y is blocked by A if
and only if every simple chain between X and Y is blocked by A.
Example 2.2 Consider the DAG in Figure 2.5.
1. X and R are d-separated by {Y, Z} because the chain [X, Y, R] is blocked
at Y , and the chain [X, Z, R] is blocked at Z.
2. X and T are d-separated by {Y, Z} because the chain [X, Y, R, T ] is blocked
at Y , the chain [X, Z, R, T ] is blocked at Z, and the chain [X, Z, S, R, T ]
is blocked at Z and at S.
3. W and T are d-separated by {R} because the chains [W, Y, R, T ] and
[W, Y, X, Z, R, T ] are both blocked at R.
4. Y and Z are d-separated by {X} because the chain [Y, X, Z] is blocked at
X, the chain [Y, R, Z] is blocked at R, and the chain [Y, R, S, Z] is blocked
at S.
5. W and S are d-separated by {R, Z} because the chain [W, Y, R, S] is blocked
at R, the chains [W, Y, R, Z, S] and [W, Y, X, Z, S] are both blocked at Z.
6. W and S are also d-separated by {Y, Z} because the chain [W, Y, R, S] is
blocked at Y , the chain [W, Y, R, Z, S] is blocked at Y , R, and Z, and the
chain [W, Y, X, Z, S] is blocked at Z.
7. W and S are also d-separated by {Y, X}. You should determine why.
8. W and X are d-separated by ∅ because the chain [W, Y, X] is blocked at Y ,
the chain [W, Y, R, Z, X] is blocked at R, and the chain [W, Y, R, S, Z, X]
is blocked at S.
9. W and X are not d-separated by {Y } because the chain [W, Y, X] is not
blocked at Y since Y ²{Y } and clearly it could not be blocked anywhere else.
10. W and T are not d-separated by {Y } because, even though the chain
[W, Y, R, T ] is blocked at Y , the chain [W, Y, X, Z, R, T ] is not blocked at Y
since Y ²{Y } and this chain is not blocked anywhere else because no other
nodes are in {Y } and there are no other head-to-head meetings on it.
Definition 2.4 Let G = (V, E) be a DAG, and A, B, and C be mutually disjoint
subsets of V. We say A and B are d-separated by C in G if for every X ∈ A and
Y ∈ B, X and Y are d-separated by C. We write
IG (A, B|C).
If C = ∅, we write only
IG (A, B).
74
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
Example 2.3 Consider the DAG in Figure 2.5. We have
IG ({W, X}, {S, T }|{R, Z})
because every chain between W and S, W and T , X and S, and X and T is
blocked by {R, Z}.
We write IG (A, B|C) because, as we show next, d-separation identifies all and
only those conditional independencies entailed by the Markov condition for G.
We need the following three lemmas to prove this:
Lemma 2.1 Let P be a probability distribution of the variables in V and G =
(V, E) be a DAG. Then (G, P ) satisfies the Markov condition if and only if for
every three mutually disjoint subsets A, B, C ⊆ V, whenever A and B are dseparated by C, A and B are conditionally independent in P given C. That is,
(G, P ) satisfies the Markov condition if and only if
IG (A, B|C) =⇒ IP (A, B|C).
(2.1)
Proof. The proof that, if (G, P ) satisfies the Markov condition, then each dseparation implies the corresponding conditional independency is quite lengthy
and can be found in [Verma and Pearl, 1990] and in [Neapolitan, 1990].
As to the other direction, suppose each d-separation implies a conditional
independency. That is, suppose Implication 2.1 holds. It is not hard to see that
a node’s parents d-separate the node from all its nondescendents that are not its
parents. That is, if we denote the sets of parents and nondescendents of X by
PAX and NDX respectively, we have
IG ({X}, NDX − PAX |PAX ).
Since Implication 2.1 holds, we can therefore conclude
IP ({X}, NDX − PAX |PAX ),
which clearly states the same conditional independencies as
IP ({X}, NDX |PAX ),
which means the Markov condition is satisfied.
According to the previous lemma, if A and B are d-separated by C in G,
the Markov condition entails IP (A, B|C). For this reason, if (G, P ) satisfies the
Markov condition, we say G is an independence map of P .
We close with an intuitive explanation for why every d-separation is a conditional independency. If G = (V, E) and (G, P ) satisfies the Markov condition,
any dependency in P between two variables in V would have to be through
a chain between them in G that has no head-to-head meetings. For example,
suppose P satisfies the Markov condition with the DAG in Figure 2.5. Any
2.1. ENTAILED CONDITIONAL INDEPENDENCIES
75
dependency in P between X and T would have to be either through the chain
[X, Y, R, T ] or the chain [X, Z, R, T ]. There could be no dependency through the
chain [X, Z, S, R, T ] owing to the head-to-head meeting at S. If we instantiate
a variable on a chain with no head-to-head meeting, we block the dependency
through that chain. For example, if we instantiate Y we block the dependency
between X and T through the chain [X, Y, R, T ], and if we instantiate Z we
block the dependency between X and T through the chain [X, Z, R, T ]. If we
block all such dependencies, we render the two variables independent. For example, the instantiation of Y and Z render X and T independent. In summary,
the fact that we have IG ({X}, {T }|{Y, Z}) means we have IP ({X}, {T }|{Y, Z}).
If every chain between two nodes contains a head-to-head meeting, there is no
chain through which they could be dependent, and they are independent. For
example, if P satisfies the Markov condition with the DAG in Figure 2.5, W and
X are independent in P . That is, the fact that we have IG ({W }, {X}) means
we have IP ({W }, {X}). Note that we cannot conclude IP ({W }, {X}|{Y }) from
the Markov condition, and we do not have IG ({W }, {X}|{Y }).
Every Entailed Conditional Independency is Identified by d-separation
Could there be conditional independencies, other than those identified by dseparation, that are entailed by the Markov condition? The answer is no. The
next two lemmas prove this. First we have a definition.
Definition 2.5 Let V be a set of random variables, and A1 , B1 , C1 , A2 ,B2 , and
C2 be subsets of V. We say conditional independency IP (A1 , B1 |C1 ) is equivalent to conditional independency IP (A2 , B2 |C2 ) if for every probability distribution P of V, IP (A1 , B1 |C1 ) holds if and only if IP (A2 , B2 |C2 ) holds.
Lemma 2.2 Any conditional independency entailed by a DAG, based on the
Markov condition, is equivalent to a conditional independency among disjoint
sets of random variables.
Proof. The proof is developed in Exercise 2.4.
Due to the preceding lemma, we need only discuss disjoint sets of random
variables when investigating conditional independencies entailed by the Markov
condition. The next lemma states that the only such conditional independencies
are those that correspond to d-separations:
Lemma 2.3 Let G = (V, E) be a DAG, and P be the set of all probability
distributions P such that (G, P ) satisfies the Markov condition. Then for every
three mutually disjoint subsets A, B, C ⊆ V,
IP (A, B|C) for all P ∈ P =⇒ IG (A, B|C).
Proof. The proof can be found in [Geiger and Pearl, 1990].
Before stating the main theorem concerning d-separation, we need the following definition:
76
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
X
Y
Z
P(x1) = a
P(x2) = 1-a
P(y1|x1) = 1 - (b + c)
P(y2|x1) = c
P(y3|x1) = b
P(z1|y1) = e
P(z2|y1) = 1 - e
P(y1|x2) = 1 - (b + d)
P(y2|x2) = d
P(y3|x2) = b
P(z1|y2) = e
P(z2|y2) = 1 - e
P(z1|y3) = f
P(z2|y3) = 1 - f
Figure 2.6: For this (G, P ), we have IP ({X}, {Z}) but not IG ({X}, {Z}).
Definition 2.6 We say conditional independency IP (A, B|C) is identified by
d-separation in G if one of the following holds:
1. IG (A, B|C).
2. A, B, and C are not mutually disjoint; A0 , B0 , and C0 are mutually disjoint,
IP (A, B|C) and IP (A0 , B0 |C0 ) are equivalent, and we have IG (A0 , B0 |C0 ).
Theorem 2.1 Based on the Markov condition, a DAG G entails all and only
those conditional independencies that are identified by d-separation in G.
Proof. The proof follows immediately from the preceding three lemmas.
You must be careful to interpret Theorem 2.1 correctly. A particular distribution P , that satisfies the Markov condition with G, may have conditional
independencies that are not identified by d-separation. For example, consider
the Bayesian network in Figure 2.6. It is left as an exercise to show IP ({X}, {Z})
for the distribution P in that network. Clearly, IG ({X}, {Z}) is not the case.
However, there are many distributions, which satisfy the Markov condition with
the DAG in that figure, that do not have this independency. One such distribution is the one given in Example 1.25 (with X, Y , and Z replaced by V ,
C, and S respectively). The only independency, that exists in all distributions satisfying the Markov condition with this DAG, is IP ({X}, {Z}|{Y }), and
IG ({X}, {Z}|{Y }) is the case.
2.1.3
Finding d-Separations
Since d-separations entail conditional independencies, we want an efficient algorithm for determining whether two sets are d-separated by another set. We
develop such an algorithm next. After that, we show a useful application of the
algorithm.
2.1. ENTAILED CONDITIONAL INDEPENDENCIES
Y
77
2
3
V
1
Q
Z
5
X
1
W
U
S
2
4
3
M
N
T
Figure 2.7: If the set of legal pairs is {(X → Y, Y → V ), (Y → V, V → Q),
(X → W, W → S), (X → U, U → T ), (U → T, T → M ), (T → M, M → S),
(M → S, S → V ), (S → V, V → Q)}, and we are looking for the nodes reachable
from {X}, Algorithm 2.1 labels the edges as shown. Reachable nodes are shaded.
An Algorithm for Finding d-Separations
We will develop an algorithm that finds the set of all nodes d-separated from
one set of nodes B by another set of nodes A. To accomplish this, we will first
find every node X such that there is at least one active chain given A between X
and a node in B. This latter task can be accomplished by solving the following
more general problem first. Suppose we have a directed graph (not necessarily
acyclic), and we say that certain edges cannot appear consecutively in our paths
of interest. That is, we identify certain ordered pairs of edges (U → V, V → W )
as legal and the remaining as illegal. We call a path legal if it does not contain
any illegal ordered pairs of edges, and we say Y is reachable from X if there
is a legal path from X to Y . Note that we are looking only for paths; we are
not looking for chains that are not paths. We can find the set R of all nodes
reachable from X as follows: We note that any node V such that the edge
X → V exists is reachable. We label each such edge with a 1, and add each
such V to R. Next for each such V , we check all unlabeled edges V → W and
see if (X → V, V → W ) is a legal pair. We label each such edge with a 2 and
we add each such W to R. We then repeat this procedure with V taking the
place of X and W taking the place of V . This time we label the edges found
with a 3. We keep going in this fashion until we find no more legal pairs. This
is similar to a breadth-first graph search except we are visiting links rather than
nodes. In this way, we may investigate a given node more than once. Of course,
we want to do this because there may be a legal path through a given node
even though another edge reaches a dead-end at the node. Figure 2.7 illustrates
this method. The algorithm that follows, which is based on an algorithm in
[Geiger et al, 1990a], implements it.
78
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
Before giving the algorithm, we discuss how we present algorithms. We
use a very loose C++ like pseudocode. That is, we use a good deal of simple
English description, we ignore restrictions of the C++ language such as the
inability to declare local arrays, and we freely use data types peculiar to the
given application without defining them. Finally, when it will only clutter rather
than elucidate the algorithm, we do not define variables. Our purpose is to
present the algorithm using familiar, clear control structures rather than adhere
to the dictates of a programming language.
Algorithm 2.1 Find Reachable Nodes
Problem: Given a directed graph and a set of legal ordered pairs of edges,
determine the set of all nodes reachable from a given set of nodes.
Inputs: a directed graph G = (V, E), a subset B ⊂ V, and a rule for determining whether two consecutive edges are legal.
Outputs: the subset R ⊂ V of all nodes reachable from B.
void f ind_reachable_nodes (directed_graph G = (V, E),
set-of-nodes B,
set-of-nodes& R)
{
for (each X ∈ B) {
add X to R;
for (each V such that the edge X → V exists) {
add V to R;
label X → V with 1;
}
}
i = 1;
f ound = true;
while (f ound) {
found = false;
for (each V such that U → V is labeled i)
for (each unlabeled edge V → W
such that (U → V ,V → W ) is legal) {
add W to R;
label V → W with i + 1;
found = true;
}
i = i + 1;
}
}
2.1. ENTAILED CONDITIONAL INDEPENDENCIES
79
Geiger at al [1990b] proved Algorithm 2.1 is correct. We analyze it next.
Analysis of Algorithm 2.1 (Find Reachable Nodes)
Let n be the number of nodes and m be the number of edges. In
the worst case, each of the nodes can be reached from n entry points
(Note that the graph is not necessarily a DAG; so there can be
edge from a node to itself.). Each time a node is reached, an edge
emanating from it may need to be re-examined. For example, in
Figure 2.7 the edge S → V is examined twice. This means each
edge may be examined n times, which implies the worst-case time
complexity is the following:
W (m, n) ∈ θ(mn).
Next we address the problem of identifying the set of nodes D that are dseparated from B by A in a DAG G = (V, E). First we will find the set R such
that Y ∈ R if and only if either Y ∈ B or there is at least one active chain given
A between Y and a node in B. Once we find R, D = V − (A ∪ R).
If there is an active chain ρ between node X and some other node, then
every 3-node subchain U − V − W of ρ has the following property: Either
1. U − V − W is not head-to-head at V and V is not in A; or
2. U − V − W is head-to-head at V and V is or has a descendent in A.
Initially we may try to mimic Algorithm 2.1. We say we are mimicking Algorithm
2.1 because now we are looking for chains that satisfy certain conditions; we are
not restricting ourselves to paths as Algorithm 2.1 does. We mimic Algorithm
2.1 as follows: We call a pair of adjacent links (U − V ,V − W ) legal if and
only if U − V − W satisfies one of the two conditions above. Then we proceed
from X as in Algorithm 2.1 numbering links and adding reachable nodes to R.
This method finds only nodes that have an active chain between them and X,
but it does not always find all of them. Consider the DAG in Figure 2.8 (a).
Given A is the only node in A and X is the only edge in B, the edges in that
DAG are numbered according to the method just described. The active chain
X → A ← Z ← T ← Y is missed because the edge T → Z is already numbered
by the time the chain A ← Z ← T is investigated, which means the chain
Z ← T ← Y is never investigated. Since this is the only active chain between
X and Y , Y is not be added to R.
We can solve this problem by creating from G = (V, E) a new directed graph
G0 = (V, E0 ), which has the links in G going in both directions. That is,
E0 = E ∪ {U → V such that V → U ∈ E}.
We then apply Algorithm 2.1 to G0 calling (U → V ,V → W ) legal in G0 if
and only if U − V − W satisfies one of the two conditions above in G. In this
80
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
Y
Y
4
X
1
T
2
Z
X
1
T
2
Z
3
1
2
1
2
A
A
(a)
(b)
Figure 2.8: The directed graph G0 in (b) is created from the DAG G in (a) by
making each link go in both directions. The numbering of the edges in (a) is
the result of applying a mimic of Algorithm 2.1 to G, while the numbering of
the edges in (b) is the result of applying Algorithm 2.1 to G0 .
way every active chain between X and Y in G has associated with it a legal
path from X to Y in G0 , and will therefore not be missed. Figure 2.8 (b)
shows G0 , when G is the DAG in Figure 2.8 (a), along with the edges numbered
according to this application of Algorithm 2.1. The following algorithm, taken
from [Geiger et al, 1990a], implements the method.
Algorithm 2.2 Find d-Separations
Problem: Given a DAG, determine the set of all nodes d-separated from one
set of nodes by another set of nodes.
Inputs: a DAG G = (V, E) and two disjoint subsets A, B ⊂ V.
Outputs: the subset D ⊂ V containing all nodes d-separated from every node
in B by A. That is, IG (B, D|A) holds and no superset of D has this property.
void f ind_d_separations (DAG G = (V, E),
set-of-nodes A, B,
set-of-nodes& D)
{
DAG G0 = (V, E0 );
2.1. ENTAILED CONDITIONAL INDEPENDENCIES
}
for (each V ∈ V) {
if (V ∈ A)
in[V ] = true;
else
in[V ] = false;
if (V is or has a descendent in A)
descendent[V ] = true;
else
descendent[V ] = false;
}
E0 = E ∪ {U → V such that V → U ∈ E};
// Call Algorithm 2.1 as follows:
f ind_reachable_nodes(G0 = (V, E0 ), B, R);
// Use this rule to decide whether (U → V, V → W ) is legal in G0 :
// The pair (U → V, V → W ) is legal if and only if U 6= W
// and one of the following hold:
// 1) U − V − W is not head-to-head in G and in[V ] is false;
// 2) U − V − W is head-to-head in G and descendent[V ] is true.
D = V − (A ∪ R);
// We do not need to remove B because B ⊆ R.
Next we analyze the algorithm:
Analysis of Algorithm 2.2 (Find d-Separations)
Although Algorithm 2.1’s worst case time complexity is in θ(mn),
where n is the number of nodes and m is the number of edges,
we will show this application of it requires only θ(m) time in the
worst case. We can implement the construction of descendent[V ]
as follows. Initially set descendent[V ] = true for all nodes in A.
Then follow the incoming edges in A to their parents, their parents’
parents, and so on, setting descendent[V ] = true for each node found
along the way. In this way, each edge is examined at most once, and
so the construction requires θ(m) time. Similarly, we can construct
in[V ] in θ(m) time.
Next we show that the execution of Algorithm 2.1 can also be done
in θ(m) time (assuming m ≥ n). To accomplish this, we use the
following data structure to represent G. For each node we store a
list of the nodes that point to that node. For example, this list for
node T in Figure 2.8 (a) is {X, Y }. Call this list the node’s inlist.
We then create an outlist for each node, which contains all the
node’s to which a node points. For example, this list for node X in
Figure 2.8 (a) is {A, T }. Clearly, these lists can be created from the
81
82
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
inlists in θ(m) time. Now suppose Algorithm 2.1 is currently trying
to determine for edge U → V in G0 which pairs (U → V, V → W )
are legal. We simply choose all the nodes in V ’s inlist or outlist or
both according to the following pseudocode:
if (U → V in G) {
// U points to V in G.
if (descendent[V ] == true)
choose all nodes W in V ’s inlist;
if (in[V ] == false)
choose all nodes W in V ’s outlist;
}
else {
// V points to U in G.
if (in[V ] == true)
choose no nodes;
else choose all nodes W in V ’s inlist and in V ’s outlist;
}
So for each edge U → V in G0 we can find all legal pairs (U →
V, V → W ) in constant time. Since Algorithm 2.1 only looks for
these legal pairs at most once for each edge U → V , the algorithm
runs in θ(m) time.
Next we prove the algorithm is correct.
Theorem 2.2 The set D returned by Algorithm 2.2 contains all and only nodes
d-separated from every node in B by A. That is, we have IG (B, D|A) and no
superset of D has this property.
Proof. The set R determined by the algorithm contains all nodes in B (because
Algorithm 2.1 initially adds nodes in B to R) and all nodes reachable from B
/ A ∪ B, the chain
via a legal path in G0 . For any two nodes X ∈ B and Y ∈
X − · · · − Y is active in G if and only if the path X → · · · → Y is legal in G0 .
Thus R contains the nodes in B plus all and only those nodes that have active
chains between them and a node in B. By the definition of d-separation, a node
is d-separated from every node in B by A if the node is not in A ∪ B and there
is no active chain between the node and a node in B. Thus D = V − (A ∪ R) is
the set of all nodes d-separated from every node in B by A.
An Application
In general, the inference problem in Bayesian networks is to determine P (B|A),
where A and B are two sets of variables. In the application of Bayesian networks
to decision theory, which is discussed in Chapter 5, we are often interested in
determining how sensitive our decision is to each parameter in the network
so that we do not waste effort trying to refine values which do not affect the
decision. This matter is discussed more in [Shachter, 1988]. Next we show how
2.1. ENTAILED CONDITIONAL INDEPENDENCIES
83
PX
X
P(x1| px) = px
P(x2| px) = 1-px
Figure 2.9: PX is a variable whose possible values are the probabilities we may
assign to x1.
H
B
L
F
C
Figure 2.10: A DAG.
Algorithm 2.2 can be used to determine which parameters are irrelevant to a
given computation.
Suppose variable X has two possible value x1 and x2, and we have not yet
ascertained P (x). We can create a variable PX whose possible values lie in the
interval [0, 1], and represent P (X = x) using the Bayesian network in Figure 2.9.
In Chapter 6 we will discuss assigning probabilities to the possible values of Px
in the case where the probabilities are relative frequencies. In general, we can
represent the possible values of the parameters in the conditional distributions
associated with a node using a set of auxiliary parent nodes. Figure 2.11 shows
one such parent node for each node in the DAG in Figure 2.10. In general, each
node can have more than one auxiliary parent node, and each auxiliary parent
node can represent a set of random variables. However, this is not important to
our present discussion; so we show only one node representing a single variable
for the sake of simplicity. You are referred to Chapters 6 and 7 for the details
of this representation. Let G00 be the DAG obtained from G by adding these
auxiliary parent nodes, and let P be the set of auxiliary parent nodes. Then to
determine which parameters are necessary to the calculation of P (B|A) in G,
we need only first use Algorithm 2.1 to determine D such that IG00 (B, D|A) and
no superset of D has this property, and then take D ∩ P.
84
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
PH
PB
H
B
PF
PL
L
F
PX
X
Figure 2.11: Each shaded node is an auxiliary parent node representing possible
values of the parameters in the conditional distributions of the child.
Example 2.4 Let G be the DAG in Figure 2.10. Then G00 is as shown in Figure
2.11. To determine P (f) we need ascertain all and only the values of PH , PB ,
PL , and PF because we have IG00 ({F }, {PX }), and PX is the only auxiliary
parent variable d-separated from {F } by the empty set. To determine P (f |b)
we need ascertain all and only the values of PH , PL , and PF because we have
IG00 ({F }, {PB , PX }|{B}), and PB and PX are the only auxiliary parent variables
d-separated from {F } by {B}. To determine P (f |b, x) we need ascertain all and
only the values of PH , PL , PF , and PX , because IG00 ({F }, {PB }|{B, X}), and
PB is the only auxiliary parent variables d-separated from {F } by {B, X}.
It is left as an exercise to write an algorithm implementing the method just
described.
2.2
Markov Equivalence
Many DAGs are equivalent in the sense that they have the same d-separations.
For example, each of the DAGs in Figure 2.12 has the d-separations IG ({Y }, {Z}|
{X}) and IG ({X}, {W }| {Y, Z}), and these are the only d-separations each has.
After stating a formal definition of this equivalence, we give a theorem showing
how it relates to probability distributions. Finally, we establish a criterion for
recognizing this equivalence.
Definition 2.7 Let G1 = (V, E1 ) and G2 = (V, E2 ) be two DAGs containing
the same set of variables V. Then G1 and G2 are called Markov equivalent
2.2. MARKOV EQUIVALENCE
X
Y
X
X
Z
W
85
Y
Z
W
Y
Z
W
Figure 2.12: These DAGs are Markov equivalent, and there are no other DAGs
Markov equivalent to them.
if for every three mutually disjoint subsets A, B, C ⊆ V, A and B are d-separated
by C in G1 if and only if A and B are d-separated by C in G2 . That is
IG1 (A, B|C) ⇐⇒ IG2 (A, B|C).
Although the previous definition has only to do with graph properties, its
application is in probability due to the following theorem:
Theorem 2.3 Two DAGs are Markov equivalent if and only if, based on the
Markov condition, they entail the same conditional independencies.
Proof. The proof follows immediately from Theorem 2.1.
Corollary 2.1 Let G1 = (V, E1 ) and G2 = (V, E2 ) be two DAGs containing the
same set of variables V. Then G1 and G2 are Markov equivalent if and only if
for every probability distribution P of V, (G1 , P ) satisfies the Markov condition
if and only if (G2 , P ) satisfies the Markov condition.
Proof. The proof is left as an exercise.
Next we develop a theorem that shows how to identify Markov equivalence.
Its proof requires the following three lemmas:
Lemma 2.4 Let G = (V, E) be a DAG and X, Y ∈ V. Then X and Y are
adjacent in G if and only if they are not d-separated by some set in G.
Proof. Clearly, if X and Y are adjacent, no set d-separates them as no set can
block the chain consisting of the edge between them.
In the other direction, suppose X and Y are not adjacent. Either there is no
path from X to Y or there is no path from Y to X for otherwise we would have
a cycle. Without loss of generality, assume there is no path from Y to X. We
will show that X and Y are d-separated by the set PAY consisting of all parents
of Y . Clearly, any chain ρ between X and Y , such that the edge incident to Y
86
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
has its head at Y , is blocked by PAY . Consider any chain ρ between X and Y
such that the edge incident to Y has its tail at Y . There must be a head-to-head
meeting on ρ because otherwise it would be path from Y to X. Consider the
head-to-head node Z closest to Y on ρ. The node Z cannot be a parent of Y
because otherwise we would have a cycle. This implies ρ is blocked by PAY ,
which completes the proof.
Corollary 2.2 Let G = (V, E) be a DAG and X, Y ∈ V. Then if X and Y are
d-separated by some set, they are d-separated either by the set consisting of the
parents of X or the set consisting of the parents of Y .
Proof. The proof follows from the proof of Lemma 2.4.
Lemma 2.5 Suppose we have a DAG G = (V, E) and an uncoupled meeting
X − Z − Y . Then the following are equivalent:
1. X − Z − Y is a head-to-head meeting.
2. There exists a set not containing Z that d-separates X and Y .
3. All sets containing Z do not d-separate X and Y .
Proof. We will show 1 ⇒ 2 ⇒ 3 ⇒ 1.
Show 1 ⇒ 2: Suppose X − Z − Y is a head-to-head meeting. Since X and Y
are not adjacent, Lemma 2.4 says some set d-separates them. If it contained Z,
it would not block the chain X − Z − Y , which means it would not d—separate
X and Y . So it does not contain Z.
Show 2 ⇒ 3: Suppose there exists a set A not containing Z that d-separates
X and Y . Then the meeting X − Z − Y must be head-to-head because otherwise
the chain X − Z − Y would not be blocked by A. However, this means any set
containing Z does not block X − Z − Y and therefore does not d-separate X and
Y.
Show 3 ⇒ 1: Suppose X − Z − Y is not a head-to-head meeting. Since X
and Y are not adjacent, Lemma 2.4 says some set d-separates them. That set
must contain Z because it must block X − Z − Y . So it is not the case that all
sets containing Z do not d-separate X and Y .
Lemma 2.6 If G1 and G2 are Markov equivalent, then X and Y are adjacent
in G1 if and only if they are adjacent in G2 . That is, Markov equivalent DAGs
have the same links (edges without regard for direction).
Proof. Suppose X and Y are adjacent in G1 . Lemma 2.4 implies they are
not d-separated in G1 by any set. Since G1 and G2 are Markov equivalent, this
means they are not d-separated in G2 by any set. Lemma 2.4 therefore implies
they are adjacent in G2 . Clearly, we have the same proof with the roles of G1
and G2 reversed. This proves the lemma.
We now give the theorem that identifies Markov equivalence. This theorem
was first stated in [Pearl et al, 1989].
2.2. MARKOV EQUIVALENCE
87
Theorem 2.4 Two DAGs G1 and G2 are Markov equivalent if and only if they
have the same links (edges without regard for direction) and the same set of
uncoupled head-to-head meetings.
Proof. Suppose the DAGs are Markov equivalent. Lemma 2.6 says they have
the same links. Suppose there is an uncoupled head-to-head meeting X → Z ←
Y in G1 . Lemma 2.5 says there is a set not containing Z that d-separates X
and Y in G1 . Since G1 and G2 are Markov equivalent, this means there is a set
not containing Z that d-separates X and Y in G2 . Again applying Lemma 2.5,
we conclude X − Z − Y is an uncoupled head-to-head meeting in G2 .
In the other direction, suppose two DAGs G1 and G2 have the same links
and the same set of uncoupled head-to-head meetings. The DAGs are equivalent
if two nodes X and Y are not d-separated in G1 by some set A ⊂ V if and only
if they are not d-separated in G2 by A. Without loss of generality, we need only
show this implication holds in one direction because the same proof can be used
to go in the other direction. If X and Y are not d-separated in G1 by A, then
there is at least one active chain (given A) between X and Y in G1 . If there
is an active chain between X and Y in G2 , then X and Y are not d-separated
in G2 by A. So we need only show the existence of an active chain between X
and Y in G1 implies the existence of an active chain between X and Y in G2 .
To that end, let N = V − A, label all nodes in N with an N , let ρ1 be an active
chain in G1 , and let ρ2 be the chain in G2 consisting of the same links. If ρ2
is not active, we will show that we can create a shorter active chain between X
and Y in G1 . In this way, we can keep creating shorter active chains between
X and Y in G1 until the corresponding chain in G2 is active, or until we create
a chain with no intermediate nodes between X and Y in G1 . In this latter case,
X and Y are adjacent in both DAGs, and the direct link between them is our
desired active chain in G2 . Assuming ρ2 is not active, we have two cases:
Case 1: There is at least one node A ∈ A responsible for ρ2 being blocked. That
is, there is a head-to-tail or tail-to-tail meeting at A on ρ2 . There must be a
head-to-head meeting at A on ρ1 because otherwise ρ1 would be blocked. Since
we’ve assumed the DAGs have the same set of uncoupled head-to-head meetings,
this means there must be an edge connecting the nodes adjacent to A in the
chains. Furthermore, these nodes must be in N because there is not a head-tohead meeting at either of them on ρ1 . This is depicted in Figure 2.13 (a). By
way of induction, assume we have sets of consecutive nodes in N on the chains
on both sides of A, the nodes all point towards A on ρ1 , and there is an edge
connecting the far two nodes N 0 and N 00 in these sets. This situation is depicted
in Figure 2.13 (b). Consider the chain σ1 in G1 between X and Y obtained by
using this edge to take a shortcut N 0 –N 00 in ρ1 around A. If there is not a
head-to-head meeting on σ1 at N 0 (Note that this includes the case where N 0 is
X.), σ1 is not blocked at N 0 . Similarly, if there is not a head-to-head meeting on
σ1 at N 00 , σ1 is not blocked at N 00 . If σ 1 is not blocked at N 0 or N 00 , we are done
because σ 1 is our desired shorter active chain. Suppose there is a head-to-head
meeting at one of them in σ1 . Clearly, this could happen at most at one of them.
88
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
Without loss of generality, say it is at N 00 . This implies N 00 6= Y , which means
there is a node to the right (closer to Y ) on the chain. Consider the chain σ2
in G2 consisting of the same links as σ 1 . There are two cases:
1. There is not a head-to-head meeting on σ2 at N 00 . Consider the node
to the right of N 00 on the chains. This node cannot be in A because it
points towards N 00 on ρ1 . We have therefore created a new instance of
the situation depicted in Figure 2.13 (b), and in this instance the node
corresponding to N 00 is closer on ρ1 to Y . This is depicted in Figure 2.13
(c). Inductively, we must therefore eventually arrive at an instance where
either 1) there is not a head-to-head meeting at either side in G1 (that is,
at the nodes corresponding to N 0 and N 00 on the chain corresponding to
σ1 ). This would at least happen when we reached both X and Y ; or 2)
there are head-to-head meetings on the same side in both G1 and G2 . In
the former situation we have found our shorter active path in G1 , and in
the latter we have the second case:
2. There is also a head-to-head meeting on σ 2 at N 00 . It is left as an exercise to show that in this case there must be a head-to-head meeting at a
node N ∗ ∈ N somewhere between N 0 and N 00 (including N 00 ) on ρ2 , and
there cannot be a head-to-head meeting at N ∗ on ρ1 (Recall and ρ1 is not
blocked.). Therefore, there must be an edge connecting the nodes on either
side of N ∗ . Without loss of generality, assume N ∗ is between A and Y .
The situation is then as depicted in Figure 2.13 (d). We have not labeled
the node to the left of N ∗ because it could be but is not necessarily A. The
direction of the edge connecting the nodes on either side of N ∗ on ρ1 must
be towards A because otherwise we would have a cycle. When we take a
shortcut around N ∗ , the node on N ∗ ’s right still has an edge leaving it
from the left and the node on N ∗ ’s left still has an edge coming into it
from the right. So this shortcut cannot be blocked in G1 at either of these
nodes. Therefore, this shortcut must result in a shorter active chain in
G1 .
Case 2: There are no nodes in A responsible for ρ2 being blocked. Then there
must be at least one node N 0 ∈ N responsible for ρ2 being blocked, which means
there must be a head-to-head meeting on ρ2 at N 0 . Since ρ1 is not blocked,
there is not a head-to-head meeting on ρ1 at N 0 . Since we’ve assumed the two
DAGs have the same set of uncoupled head-to-head meetings, this means the
nodes adjacent to N 0 on the chains are adjacent to each other. Since there is a
head-to-head meeting on ρ2 at N 0 , there cannot be a head—to-head meeting on ρ2
at either of these nodes (the ones adjacent to N 0 on the chains). These nodes
therefore cannot be in A because we’ve assumed no nodes in A are responsible
for ρ2 being blocked. Since ρ1 is not blocked, we cannot have a head-to-head
meeting on ρ1 at a node in N. Therefore, the only two possibilities (aside from
symmetrical ones) in G1 are the ones depicted in Figures 2.14 (a) and (b).
Clearly, in either case by taking the shortcut around N 0 , we have a shorter
active chain in G1 .
2.2. MARKOV EQUIVALENCE
89
D1
X
N
A
N
Y
D2
X
N
A
N
Y
(a)
D1
X
N'
N
A
N
N''
Y
D2
X
N'
N
A
N
N''
Y
(b)
D1
X
N'
N
A
N
N''
N
Y
D2
X
N'
N
A
N
N''
N
Y
(c)
D1
X
N'
N
A
N
N*
N
N''
Y
D2
X
N'
N
A
N
N*
N
N''
Y
(d)
Figure 2.13: The figure used to prove Case 1 in Theorem 2.4.
N
N'
(a)
N
N
N'
N
(b)
Figure 2.14: In either case, taking the shortcut around N 0 results in a shorter
active chain in G1 .
90
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
W
X
W
Y
X
Z
R
Y
Z
S
R
S
(a)
(b)
W
W
X
Y
X
Z
R
Z
S
(c)
Y
R
S
(d)
Figure 2.15: The DAGs in (a) and (b) are Markov equivalent. The DAGs in (c)
and (d) are not Markov equivalent to the first two DAGs or to each other.
2.2. MARKOV EQUIVALENCE
91
Example 2.5 The DAGs in Figure 2.15 (a) and (b) are Markov equivalent
because they have the same links and the only uncoupled head-to-head meeting
in both is X → Z ← Y . The DAG in Figure 2.15 (c) is not Markov equivalent
to the first two because it has the link W − Y . The DAG in Figure 2.15 (d) is
not Markov equivalent to the first two because, although it has the same links,
it does not have the uncoupled head-to-head meeting X → Z ← Y . Clearly, the
DAGs in Figure 2.15 (c) and (d) are not Markov equivalent to each other either.
It is straightforward that Theorem 2.4 enables us to develop a polynomialtime algorithm for determining whether two DAGs are Markov equivalent. We
simply check if they have the same links and uncoupled head-to-head meetings.
It is left as an exercise to write such an algorithm.
Furthermore, Theorem 2.4 gives us a simple way to represent a Markov
equivalence class with a single graph. That is, we can represent a Markov
equivalent class with a graph that has the same links and the same uncoupled
head-to-head meeting as the DAGs in the class. Any assignment of directions to
the undirected edges in this graph, that does not create a new uncoupled headto-head meeting or a directed cycle, yields a member of the equivalence class.
Often there are edges other than uncoupled head-to-head meetings which must
be oriented the same in Markov equivalent DAGs. For example, if all DAGs in
a given Markov equivalence class have the edge X → Y , and the uncoupled
meeting X → Y − Z is not head-to-head, then all the DAGs in the equivalence
class must have Y − Z oriented as Y → Z. So we define a DAG pattern for a
Markov equivalence class to be the graph that has the same links as the DAGs
in the equivalence class and has oriented all and only the edges common to all
of the DAGs in the equivalence class. The directed links in a DAG pattern
are called compelled edges. The DAG pattern in Figure 2.16 represents the
Markov equivalence class in Figure 2.12. The DAG pattern in Figure 2.17 (b)
represents the Markov equivalent class in Figure 2.17 (a). Notice that no DAG
Markov equivalent to each of the DAGs in Figure 2.17 (a) can have W − U
oriented as W ← U because this would create another uncoupled head-to-head
meeting.
Since all DAGs in the same Markov equivalence class have the same dseparations, we can define d-separation for DAG patterns:
Definition 2.8 Let gp be a dag pattern whose nodes are the elements of V, and
A, B, and C be mutually disjoint subsets of V. We say A and B are d-separated
by C in gp if A and B are d-separated by C in any (and therefore every) DAG
G in the Markov equivalence class represented by gp.
Example 2.6 For the DAG pattern gp in Figure 2.16 we have
Igp ({Y }, {Z}|{X})
because {Y } and {Z} are d-separated by {X} in the DAGs in Figure 2.12.
The following lemmas follow immediately from the corresponding lemmas
for DAGs:
92
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
X
Y
Z
W
Figure 2.16: This DAG pattern represents the Markov equivalence class in Figure 2.12.
Lemma 2.7 Let gp be DAG and X and Y be nodes in gp. Then X and Y are
adjacent in gp if and only if they are not d-separated by some set in gp.
Proof. The proof follows from Lemma 2.4.
Lemma 2.8 Suppose we have a DAG pattern gp and an uncoupled meeting
X − Z − Y . Then the following are equivalent:
1. X − Z − Y is a head-to-head meeting.
2. There exists a set not containing Z that d-separates X and Y .
3. All sets containing Z do not d-separate X and Y .
Proof. The proof follows from Lemma 2.5.
Owing to Corollary 2.1, if G is an independence map of a probability distribution P (i.e. (G, P ) satisfies the Markov condition), then every DAG Markov
equivalent to G is also an independence map of P . In this case, we say the DAG
pattern gp representing the equivalence class is an independence map of P .
2.3
Entailing Dependencies with a DAG
As noted at the beginning of this chapter, the Markov condition only entails
independencies; it does not entail any dependencies. As a result, many uninformative DAGs can satisfy the Markov condition with a given distribution P .
The following example illustrates this.
Example 2.7 Let Ω be the set of objects in Figure 1.2, and let P , V , S, and C
be as defined in Example 1.25. That is, P assigns a probability of 1/13 to each
object, and random variables V , S, and C are defined as follows:
2.3. ENTAILING DEPENDENCIES WITH A DAG
Z
Z
X
93
Y
X
Z
Y
X
Y
W
W
W
U
U
U
(a)
Z
X
Y
W
U
(b)
Figure 2.17: The DAG pattern in (b) represents the Markov equivalence class
in (a).
94
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
C
C
V
S
C
V
(a)
S
(b)
V
S
(c)
Figure 2.18: The probability distribution in Example 2.4 satisfies the Markov
condition with each of these DAGs.
Variable
V
S
C
Value
v1
v2
s1
s2
c1
c2
Outcomes Mapped to this Value
All objects containing a ‘1’
All objects containing a ‘2’
All square objects
All round objects
All black objects
All white objects
Then, as shown in Example 1.25, P satisfies the Markov condition with the
DAG in Figure 2.18 (a) because IP ({V }, {S}|{C}). However, P also satisfies
the Markov condition with the DAGs in Figures 2.18 (b) and (c) because the
Markov condition does not entail any independencies in the case of these DAGs.
This means that not only P but every probability distribution of V , S, and C
satisfies the Markov condition with each of these DAGs.
The DAGs in Figures 2.18 (b) and (c) are complete DAGs. Recall that a
complete DAG G = (V, E) is one in which there is an edge between every
pair of nodes. That is, for every X, Y ∈ V, either (X, Y ) ∈ E or (Y, X) ∈ E.
In general, the Markov condition entails no independencies in the case of a
complete DAG G = (V, E), which means (G, P ) satisfies the Markov condition
for every probability distribution P of the variables in V. We see then that
(G, P ) can satisfy the Markov condition without G telling us anything about P .
Given a probability distribution P of the variables in some set V and X, Y ∈
V, we say there is a direct dependency between X and Y in P if {X} and
{Y } are not conditionally independent given any subset of V. The problem
with the Markov condition alone is that it entails that the absence of an edge
between X any Y means there is no direct dependency between X any Y , but
it does not entail that the presence of an edge between X and Y means there
is a direct dependency. That is, if there is no edge between X and Y , Lemmas
2.4 and 2.1 together tell us the Markov condition entails {X} and {Y } are
conditionally independent given some set (possibly empty) of variables. For
2.3. ENTAILING DEPENDENCIES WITH A DAG
95
example, in Figure 2.18 (a), because there is no edge between V and C, we
know from Lemma 2.4 they are d-separated by some set. It turns out that set
is {C}. Lemma 2.1 therefore tells us IP ({V }, {S}|{C}). On the other hand, if
there is an edge between X and Y , the Markov condition does not entail that
{X} and {Y } are not conditionally independent given some set of variables. For
example, in Figure 2.18 (b), the edge between V and S does not mean that {V }
and {S} are not conditionally independent given some set of variables. Indeed,
we know they actually are.
2.3.1
Faithfulness
In general, we would want an edge to mean there is a direct dependency. As we
shall see, the faithfulness condition entails this. We discuss it next.
Definition 2.9 Suppose we have a joint probability distribution P of the random variables in some set V and a DAG G = (V, E). We say that (G, P ) satisfies
the faithfulness condition if, based on the Markov condition, G entails all
and only conditional independencies in P . That is, the following two conditions
hold:
1. (G, P ) satisfies the Markov condition (This means G entails only conditional independencies in P .).
2. All conditional independencies in P are entailed by G, based on the Markov
condition.
When (G, P ) satisfies the faithfulness condition, we say P and G are faithful
to each other, and we say G is a perfect map of P . When they do not, we say
they are unfaithful to each other.
Example 2.8 Let P and V , S, and C be as in Example 2.7. Then, as shown
in Example 1.25, IP ({V }, {S}|{C}), which means (G, P ) satisfies the Markov
condition if G is the DAG in Figure 1.3 (a), (b), or (c). Those DAGs are shown
again in Figure 2.19. It is left as an exercise to show that there are no other
conditional independencies in P . That is, you should show
qIP ({V }, {S})
qIP ({V }, {C})
qIP ({S}, {C}).
qIP ({V }, {C}|{S})
qIP ({C}, {S}|{V })
(It is not necessary to show, for example, qIP ({V }, {S, C}) because the first
non-independency listed above implies this one.) Therefore, (G, P ) satisfies the
faithfulness condition if G is any one of the DAGs in Figure 2.19.
The following theorems establish a criterion for recognizing faithfulness:
96
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
V
S
S
C
C
C
S
V
V
(b)
(c)
(d)
C
V
S
(a)
Figure 2.19: The probability distribution in Example 2.7 satisfies the faithfulness condition with each of the DAGs in (a), (b), and (c), and with the DAG
pattern in (d).
Theorem 2.5 Suppose we have a joint probability distribution P of the random
variables in some set V and a DAG G = (V, E). Then (G, P ) satisfies the
faithfulness condition if and only if all and only conditional independencies in
P are identified by d-separation in G.
Proof. The proof follows immediately from Theorem 2.1.
Example 2.9 Consider the Bayesian network (G, P ) in Figure 2.6, which is
shown again in Figure 2.20. As noted in the discussion following Theorem 2.1,
for that network we have IP ({X}, {Z}) but not IG ({X}, {Z}). Therefore, (G, P )
does not satisfy the faithfulness condition.
We made very specific conditional probability assignments in Figure 2.20 to
develop a distribution that is unfaithful to the DAG in that figure. If we just
arbitrarily assign conditional distributions to the variables in a DAG, are we
X
Y
Z
P(x1) = a
P(x2) = 1-a
P(y1|x1) = 1 - (b + c)
P(y2|x1) = c
P(y3|x1) = b
P(z1|y1) = e
P(z2|y1) = 1 - e
P(y1|x2) = 1 - (b + d)
P(y2|x2) = d
P(y3|x2) = b
P(z1|y2) = e
P(z2|y2) = 1 - e
P(z1|y3) = f
P(z2|y3) = 1 - f
Figure 2.20: For this (G, P ), we have IP ({X}, {Z}) but not IG ({X}, {Z}).
2.3. ENTAILING DEPENDENCIES WITH A DAG
97
likely to end up with a joint distribution that is unfaithful to the DAG? The
answer is no. A theorem to this effect in the case of linear models appears in
[Spirtes et al, 1993, 2000]. In a linear model, each variable is a linear function
of its parents and an error variable. In this case, the set of possible conditional
probability assignments to some DAG is a real space. The theorem says that
the set of all points in this space, that yield distributions unfaithful to the DAG,
form a set of Lebesgue measure zero. Intuitively, this means that almost all such
assignments yield distributions faithful to the DAG. Meek [1995a] extends this
result to the case of discrete variables.
The following theorem obtains the result that if P is faithful to some DAG,
then P is faithful to an equivalence class of DAGs:
Theorem 2.6 If (G, P ) satisfies the faithfulness condition, then P satisfies this
condition with all and only those DAGs that are Markov equivalent to G. Furthermore, if we let gp be the DAG pattern corresponding to this Markov equivalence class, the d-separations in gp identify all and only conditional independencies in P . We say that gp and P are faithful to each other, and gp is a
perfect map of P .
Proof. The proof follows immediately from Theorem 2.5.
We say a distribution P admits a faithful DAG representation if P
is faithful to some DAG (and therefore some DAG pattern). The distribution
discussed in Example 2.8 admits a faithful DAG representation. Owing to the
previous theorem, if P admits a faithful DAG representation, there is a unique
DAG pattern with which P is faithful. In general, our goal is to find that DAG
pattern whenever P admit a faithful DAG representation. Methods for doing
this are discussed in Chapters 8-11. Presently, we show not every P admits a
faithful DAG representation.
Example 2.10 Consider the Bayesian network in Figure 2.20. As mentioned
in Example 2.9, the distribution in that network has these independencies:
IP ({X}, {Z})
IP ({X}, {Z}|{Y }).
Suppose we specify values to the parameters so that these are the only independencies, and some DAG G is faithful to the distribution (Note that G is not
necessarily the DAG in Figure 2.20.). Due to Theorem 2.5, G has these and
only these d-separations:
IG ({X}, {Z})
IG ({X}, {Z}|{Y }).
Lemma 2.4 therefore implies the links in G are X − Y and Y − Z. This means
X − Y − Z is an uncoupled meeting. Since IG ({X}, {Z}), Condition (2) in
Lemma 2.5 holds. This lemma therefore implies its Condition (3) holds, which
means we cannot have IG ({X}, {Z}|{Y }). This contradiction shows there can
be no such DAG.
98
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
L
C
V
F
S
(a)
L
F
V
S
(b)
Figure 2.21: If P satisifies the faithfulness condition with the DAG in (a), the
marginal distribution of V , S, L, and F cannot satisfy the faithfulness with any
DAG. There would have to be arrows going both ways between V and S. This
is depicted in (b).
Example 2.11 Suppose we specify conditional distributions for the DAG in
Figure 2.21 (a) so that the resultant joint distribution P (v, s, c, l, f ) satisfies the
faithfulness condition with that DAG. Then the only independencies involving
only the variables V , S, L, and F are the following:
IP ({L}, {F, S})
IP ({F }, {L, V })
IP ({L}, {S})
IP ({F }, {V }).
IP ({L}, {F })
(2.2)
Consider the marginal distribution P (v, s, , l, f ) of P (v, s, c, l, f ). We will show
this distribution does not admit a faithful DAG representation. Due to Theorem
2.5, if some DAG G was faithful to that distribution, it would have these and
only these d-separations involving only the nodes V , S, L, and F :
IG ({L}, {F, S})
IG ({F }, {L, V })
IG ({L}, {S})
IG ({F }, {V }).
IG ({L}, {F })
Due to Lemma 2.4, the links in G are therefore L − V , V − S, and S − F . This
means L − V − S is an uncoupled meeting. Since IG ({L}, {S}), Lemma 2.5
therefore implies it is an uncoupled head-to-head meeting. Similarly, V − S − F
is an uncoupled head-to-head meeting. The resultant graph, which is shown in
Figure 2.21 (b), is not a DAG. This contradiction shows P (v, s, l, f ) does not
admit a faithful DAG representation. Exercise 2.20 shows an urn problem in
which four variables have this distribution.
2.3. ENTAILING DEPENDENCIES WITH A DAG
99
Pearl [1988] obtains necessary but not sufficient conditions for a probability
distribution to admit a faithful DAG representation.
Recall at the beginning of this subsection we stated that, in the case of
faithfulness, an edge between two nodes means there is a direct dependency
between the nodes. The theorem that follows obtains this result and more.
Theorem 2.7 Suppose we have a joint probability distribution P of the random
variables in some set V and a DAG G = (V, E). Then if P admits a faithful
DAG representation, gp is the DAG pattern faithful to P if and only if the
following two conditions hold:
1. X and Y are adjacent in gp if and only if there is no subset S ⊆ V such
that IP ({X}, {Y }|S). That is, X and Y are adjacent if and only if there
is a direct dependency between X and Y .
2. X − Z − Y is a head-to-head meeting in gp if and only if Z ∈ S implies
qIP ({X}, {Y }|S).
Proof. Suppose gp is the DAG pattern faithful to P . Then due to Theorem 2.6,
all and only the independencies in P are identified by d-separation in gp, which
are the d-separations in any DAG G in the equivalence class represented by gp.
Therefore, Condition 1 follows Lemma 2.4, and Condition 2 follows from and
Lemma 2.5.
In the other direction, suppose Conditions (1) and (2) hold for gp and P .
Since we’ve assumed P admits a faithful DAG representation, there is some
DAG pattern gp0 faithful to P . By what was just proved, we know Conditions (1)
and (2) also hold for gp0 and P . However, this mean any DAG G in the Markov
equivalence class represented by gp must have the same links and same set of
uncoupled head-to-head meetings as any DAG G0 in the Markov equivalence class
represented by gp0 . Theorem 2.4 therefore says G and G0 are in the same Markov
equivalence class, which means gp = gp0 .
2.3.2
Embedded Faithfulness
The distribution P (v, s, l, f ) in Example 2.11 does not admit a faithful DAG representation. However, it is the marginal of a distribution, namely P (v, s, c, l, f ),
of one which does. This is an example of embedded faithfulness, which is defined
as follows:
Definition 2.10 Let P be a joint probability distribution of the variables in V
where V ⊆ W, and G = (W, E) be a DAG. We say (G, P ) satisfies the embedded
faithfulness condition if the following two conditions hold:
1. Based on the Markov condition, G entails only conditional independencies
in P for subsets including only elements of V.
2. All conditional independencies in P are entailed by G, based on the Markov
condition.
100
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
When (G, P ) satisfies the embedded faithfulness condition, we say P is embedded faithfully in G. Notice that faithfulness is a special case of embedded
faithfulness in which W = V.
Example 2.12 Clearly, the distribution P (v, s, l, f ) in Example 2.11 is embedded faithfully in the DAG in Figure 2.21 (a).
As was done in the previous example, we often obtain embedded faithfulness by taking the marginal of a faithful distribution. The following theorem
formalizes this result:
Theorem 2.8 Let P be a joint probability distribution of the variables in W
with V ⊆ W, and G = (W, E). If (G, P ) satisfies the faithfulness condition,
and P 0 is the marginal distribution of V, then (G, P 0 ) satisfies the embedded
faithfulness condition.
Proof. The proof is obvious. .
Definition 2.10 has only to do with independencies entailed by a DAG. It
says nothing about P being a marginal of a distribution of the variables in V.
There are other cases of embedded faithfulness. Example 2.14 shows one such
case. Before giving that example, we discuss embedded faithfulness further.
The following theorems are analogous to the corresponding ones concerning
faithfulness:
Theorem 2.9 Let P be a joint probability distribution of the variables in V with
V ⊆ W, and G = (W, E). Then (G, P ) satisfies the embedded faithfulness condition if and only if all and only conditional independencies in P are identified
by d-separation in G restricted to elements of V.
Proof. The proof is left as an exercise.
Theorem 2.10 Let P be a joint probability distribution of the variables in V
with V ⊆ W, and G = (W, E). If (G, P ) satisfies the embedded faithfulness
condition, then P satisfies this condition with all those DAGs that are Markov
equivalent to G. Furthermore, if we let gp be the DAG pattern corresponding to
this Markov equivalence class, the d-separations in gp, restricted to elements of
V, identify all and only conditional independencies in P . We say P is embedded
faithfully in gp.
Proof. The proof is left as an exercise.
Note that the theorem says ‘all those DAGS’, but, unlike the corresponding
theorem for faithfulness, it does not say ‘only those DAGs’. If a distribution can
be embedded faithfully, there are an infinite number of non-Markov equivalent
DAGs in which it can be embedded faithfully. Trivially, we can always replace
an edge by a directed linked list of new variables. Figure 2.22 shows a more
complex example. The distribution P (v, s, l, f ) in Example 2.11 is embedded
faithfully in both DAGs in that figure. However, even though the DAGs contain
the same nodes, they are not Markov equivalent.
2.3. ENTAILING DEPENDENCIES WITH A DAG
L
C
101
F
Y
X
V
L
S
C
F
Y
X
V
S
Figure 2.22: Suppose the only conditional independencies in a probability distribution P of V , S, L, and F are those in Equality 2.2, which appears in
Example 2.11. Then P is embedded faithfully in both of these DAGs.
We say a probability distribution admits an embedded faithful DAG
representation if it can be embedded faithfully in some DAG. Does every
probability distribution admit an embedded faithful DAG representation? The
following example shows the answer is no.
Example 2.13 Consider the distribution in Example 2.10. Recall that it has
these and only these conditional independencies:
IP ({X}, {Z})
IP ({X}, {Z}|{Y }).
Example 2.10 showed this distribution does not admit a faithful DAG representation. We show next that it does not even admit an embedded faithful DAG
representation. Suppose it can be embedded faithfully in some DAG G. Due to
theorem 2.9, G must have these and only these d-separations among the variables
X, Y , and Z:
IG ({X}, {Z}|{Y }).
IG ({X}, {Z})
There must be a chain between X and Y with no head-to-head meetings because
otherwise we would have IG ({X}, {Y }). Similarly, there must be a chain between
Y and Z with no head-to-head meetings. Consider the resultant chain between X
and Z. If it had a head-to-head meeting at Y , it would not be blocked by {Y } because it does not have a head-to-head meeting at a node not in {Y }. This means
if it had a head-to-head meeting at Y , we would not have IG ({X}, {Z}|{Y }).
If it did not have a head-to-head meeting at Y , there would be no head-to-head
meetings on it at all, which means it would not be blocked by ∅, and we would
102
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
X
X
T
T
Y
Y
Z
Z
W
W
(a)
(b)
Figure 2.23: The DAG in (a) includes distributions of X, Y , Z, and W which
the DAG in (b) does not.
therefore not have IG ({X}, {Z}). This contradiction shows there can be no such
DAG.
We say P is included in DAG G if P is the probability distribution in a
Bayesian network containing G or P is the marginal of a probability distribution
in a Bayesian network containing G. When a probability distribution is faithful
to some DAG G, P is included in G by definition because the faithfulness
condition subsumes the Markov condition. In the case of embedded faithfulness,
things are not as simple. It is possible to embed a distribution P faithfully in
a DAG G without P being included in the DAG. The following example, taken
from [Verma and Pearl, 1991], shows such a case:
Example 2.14 Let V = {X, Y, Z, W } and W = {X, Y, Z, W, T }. The only dseparation among the variables in V in the DAGs in Figures 2.23 (a) and (b),
is IG ({Z}, {X}|{Y }). Suppose we assign conditional distributions to the DAG
in (a) so that the resultant joint distribution of W is faithful to that DAG. Then
the marginal distribution of V is faithfully embedded in both DAGs. The DAG
in (a) has the same edges as the one in (b) plus one more. So the DAG in (b)
has d-separations, (e.g. IG ({W }, {X}|{Y, T }), which the one in (a) does not
have. We will show that as a result there are distributions which are embedded
faithfully in both DAGs but are only included in the DAG in (a).
To that end, for any marginal distribution P (v) of a probability distribution
2.3. ENTAILING DEPENDENCIES WITH A DAG
103
P (w) satisfying the Markov condition with the DAG in (b), we have
X
P (x, y, z, w) =
P (w|z, t)P (z|y)P (y|x, t)P (x)P (t)
t
= P (z|y)P (x)
X
P (w|z, t)P (y|x, t)P (t).
t
Also, for any marginal distribution P (v) of a probability distribution P (w) satisfying the Markov condition with the DAGs in both figures, we have
P (x, y, z, w) = P (w|x, y, z)P (z|x, y)P (y|x)p(x)
= P (w|x, y, z)P (z|y)P (y|x)P (x).
Equating these two expressions and summing over y yields
X
X
P (w|x, y, z)P (y|x) =
P (w|z, t)P (t).
y
t
The left hand side of the previous expression contains the variable x, whereas the
right hand side does not. Therefore, for a distribution of V to be the marginal of
a distribution of W which satisfies the Markov condition with the DAG in (b),
the distribution of V must have the left hand side equal for all values of x. For
example, for all values of w and z it would need to have
X
X
P (w|x1 , y, z)P (y|x) =
P (w|x2 , y, z)P (y|x).
(2.3)
y
y
Repeating the same steps as above for the DAG in (a), we obtain that for
any marginal distribution P (v) of a probability distribution P (w) satisfying the
Markov condition with that DAG, we have
X
X
P (w|x, y, z)P (y|x) =
P (w|x, z, t)P (t).
(2.4)
y
t
Note that now the variable x appears on both sides of the equality. Suppose
we assign values to the conditional distributions in the DAG in (a) to obtain a
distribution P 0 (w) such that for some values of w and z
X
X
P 0 (w|x1 , z, t)P 0 (t) 6=
P 0 (w|x2 , z, t)P 0 (t).
t
t
Then owing to Equality 2.4 we would have for the marginal distribution P 0 (v)
X
X
P 0 (w|x1 , y, z)P 0 (y|x) 6=
P 0 (w|x2 , y, z)P 0 (y|x).
y
y
However, Equality 2.3 says these two expressions must be equal if a distribution
of V is to be the marginal of a distribution of W which satisfies the Markov
condition with the DAG in (b). So the marginal distribution P 0 (v) is not the
104
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
marginal of a distribution of W which satisfies the Markov condition with the
DAG in (b).
Suppose further that we have made conditional distribution assignments so
that P 0 (w) is faithful to the DAG (a). Then owing to the discussion at the
beginning of the example, P 0 (v) is embedded faithfully in the DAG (b). So we
have found a distribution of V which is embedded faithfully in the DAG in (b)
but is not included in it.
2.4
Minimality
Consider again the Bayesian network in Figure 2.20. The probability distribution in that network is not faithful to the DAG because it has the independency
IP ({X}, {Z}) and the DAG does not have the d-separation IG ({X}, {Z}). In
Example 2.10 we showed that it is not possible to find a DAG faithful to that
distribution. So the problem was not in our choice of DAGs. Rather it is inherent in the distribution that there is no DAG with which it is faithful. Notice
that, if we remove either of the edges from the DAG in Figure 2.20, the DAG
ceases to satisfy the Markov condition with P . For example, if we remove the
edge X → Y , we have IG ({X}, {Y, Z}) but not IP ({X}, {Y, Z}). So the DAG
does have the property that it is minimal in the sense that we cannot remove
any edges without the Markov condition ceasing to hold. Furthermore, if we add
an edge between X and Z to form a complete graph, it would not be minimal in
this sense. Formally, we have the following definition concerning the property
just discussed:
Definition 2.11 Suppose we have a joint probability distribution P of the random variables in some set V and a DAG G = (V, E). We say that (G, P ) satisfies
the minimality condition if the following two conditions hold:
1. (G, P ) satisfies the Markov condition.
2. If we remove any edges from G, the resultant DAG no longer satisfies the
Markov condition with P .
Example 2.15 Consider the distribution P in Example 2.7. The only conditional independency is IP ({V }, {S}|{C}). The DAG in Figure 2.18 (a) satisfies
the minimality condition with P because if we remove the edge C → V we have
IG ({V }, {C, S}), if we remove the edge C → S we have IG ({S}, {C, V }), and
neither of these independencies hold in P . The DAG in Figure 2.18 (b) does
not satisfy the minimality condition with P because if remove the edge V → S,
the only new d-separation is IG ({V }, {S}|{C}), and this independency does hold
in P . Finally, the DAG in Figure 2.18 (c) does satisfy the minimality condition with P because no edge can be removed without creating a d-separation that
is not an independency in P . For example, if we remove V → S, we have
IG ({V }, {S}), and this independency does not hold in P .
2.4. MINIMALITY
105
The previous example illustrates that a DAG can satisfy the minimality
condition with a distribution without being faithful to the distribution. Namely,
the only DAG in Figure 2.18 that is faithful to P is the one in (a), but the
one in (c) also satisfies the minimality condition with P . On the other hand,
the reverse is not true. Namely, a DAG cannot be faithful to a distribution
without satisfying the minimality with the distribution. The following theorem
summarizes these results:
Theorem 2.11 Suppose we have a joint probability distribution P of the random variables in some set V and a DAG G = (V, E). If (G, P ) satisfies the
faithfulness condition, then (G, P ) satisfies the minimality condition. However,
(G, P ) can satisfy the minimality condition without satisfying the faithfulness
condition.
Proof. Suppose (G, P ) satisfies the faithfulness condition and does not satisfy
the minimality condition. Since (G, P ) does not satisfy the minimality condition.
some edge (X, Y ) can be removed and the resultant DAG will still satisfy the
Markov condition with P . Due to Lemma 2.4, X and Y are d-separated by some
set in this new DAG and therefore, due to Lemma 2.1, they are conditionally
independent given this set. Since there is an edge between X and Y in G,
Lemma 2.4 implies X and Y are not d-separated by any set in G. Since (G, P )
satisfies the faithfulness condition, Theorem 2.5 therefore implies they are not
conditionally independent given any set. This contradiction proves faithfulness
implies minimality.
The probability distribution in Example 2.7 along with the DAG in Figure
2.18 (c) shows minimality does not imply faithfulness.
The following theorem shows that every probability distribution P satisfies
the minimality condition with some DAG and gives a method for constructing
one:
Theorem 2.12 Suppose we have a joint probability distribution P of the random variables in some set V. Create an arbitrary ordering of the nodes in V. For
each X ∈ V, let BX be the set of all nodes that come before X in the ordering,
and let PAX be a minimal subset of BX such that
IP ({X}, BX |PAX )
Create a DAG G by placing an edge from each node in PAX to X. Then (G, P )
satisfies the minimality condition. Furthermore, if P is strictly positive (That
is, there are no probability values equal 0.), then PAX is unique relative to the
ordering.
Proof. The proof is developed in [Pearl, 1988].
Example 2.16 Suppose V = {X, Y, Z, W } and P is a distribution that is faithful to the DAG in Figure 2.24 (a). Then Figure 2.24 (b), (c), (d), and (e) show
four DAGs satisfying the minimality condition with P obtained using the preceding theorem. The ordering used to obtain each DAG is from top to bottom
106
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
X
Y
W
Z
(a)
X
Y
X
W
Y
X
Y
Z
W
W
Z
X
Z
Z
W
Y
(b )
(c)
(d)
(e)
Figure 2.24: Four DAGs satisfying the minimality condition with P are shown
in (b), (c), (d), and (e) given that P is faithful to the DAG in (a).
2.4. MINIMALITY
107
X
X
Y
Y
Z
Z
Figure 2.25: Two minimal DAG descriptions relative to the ordering [X, Y, Z]
when P (y1|x1) = 1 and P (y2|x2) = 1.
as shown in the figure. If P is strictly positive, each of these DAGs is unique
relative to its ordering.
Notice from the previous example that a DAG satisfying the minimality
condition with a distribution is not necessarily minimal in the sense that it
contains the minimum number of edges needed to include the distribution. Of
the DAGs in Figure 2.24, only the ones in (a), (b), and (c) are minimal in this
sense. It is not hard to see that if a DAG is faithful to a distribution, then it is
minimal in this sense.
Finally, we present an example showing that the method in Theorem 2.12
does not necessarily yield a unique DAG when the distribution is not strictly
positive.
Example 2.17 Suppose V = {X, Y, Z} and P is defined as follows:
P (x1) = a
P (x2) = 1 − a
P (y1|x1) = 1
P (y2|x1) = 0
P (z1|x1) = b
P (z2|x1) = 1 − b
P (y1|x2) = 0
P (y2|x2) = 1
P (z1|x2) = c
P (z2|x2) = 1 − c
Given the ordering [X, Y, Z], both DAGs in Figure 2.25 are minimal descriptions
of P obtained using the method in Theorem 2.12.
108
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
T
S
X
Y
Z
W
Figure 2.26: If P satisfies the Markov condition with this DAG, then {T, Y, Z}
is a Markov blanket of X.
2.5
Markov Blankets and Boundaries
A Bayesian network can have a large number of nodes, and the probability of
a given node can be affected by instantiating a distant node. However, it turns
out that the instantiation of a set of close nodes can shield a node from the
affect of all other nodes. The following definition and theorem show this:
Definition 2.12 Let V be a set of random variables, P be their joint probability
distribution, and X ∈ V. Then a Markov blanket MX of X is any set of
variables such that X is conditionally independent of all the other variables
given MX . That is,
IP ({X}, V − (MX ∪ {X})|MX ).
Theorem 2.13 Suppose (G, P ) satisfies the Markov condition. Then for each
variable X, the set of all parents of X, children of X, and parents of children
of X is a Markov blanket of X.
Proof. It is straightforward that this set d-separates {X} from the set of all
other nodes in V. That is, if we call this set MX ,
IG ({X}, V − (MX ∪ {X})|MX ).
The proof therefore follows from Theorem 2.1.
Example 2.18 Suppose (G, P ) satisfies the Markov condition where G is the
DAG in Figure 2.26. Then due to Theorem 2.13 {T, Y, Z} is a Markov blanket
of X.
Example 2.19 Suppose (G, P ) satisfies the Markov condition where G is the
DAG in Figure 2.26, and (G0 , P ) also satisfies the Markov condition where G0
2.5. MARKOV BLANKETS AND BOUNDARIES
109
is the DAG G in Figure 2.26 with the edge T → X removed. Then the Markov
blanket {T, Y, Z} is not minimal in the sense that its subset {Y, Z} is also a
Markov blanket of X.
The last example motivates the following definition:
Definition 2.13 Let V be a set of random variables, P be their joint probability
distribution, and X ∈ V. Then a Markov boundary of X is any Markov
blanket such that none of its proper subsets is a Markov blanket of X.
We have the following theorem:
Theorem 2.14 Suppose (G, P ) satisfies the faithfulness condition. Then for
each variable X, the set of all parents of X, children of X, and parents of
children of X is the unique Markov boundary of X.
Proof. Let MX be the set identified in this theorem. Due to Theorem 2.13,
MX is a Markov blanket of X. Clearly there is at least one Markov boundary
for X. So if MX is not the unique Markov boundary for X, there would have to
be some other set A not equal to MX , which is a Markov boundary of X. Since
MX 6= A and MX cannot be a proper subset of A, there is some Y ∈ MX such
that Y ∈
/ A. Since A is a Markov boundary for X, we have IP ({X}, {Y }|A).
If Y is a parent or a child of X, we would not have IG ({X}, {Y }|A), which
means we would have a conditional independence which is not a d-separation.
But Theorem 2.5 says this cannot be. If Y is a parent of a child of X, let Z
be their common child. If Z ∈ A, we again would not have IG ({X}, {Y }|A). If
Z ∈
/ A, we would have IP ({X}, {Z}|A) because A is a Markov boundary of X,
but we do not have IG ({X}, {Z}|A) because X is a parent of Z. So again we
would have a conditional independence which is not a d-separation. This proves
there can be no such set A.
Example 2.20 Suppose (G, P ) satisfies the faithfulness condition where G is
the DAG in Figure 2.26. Then due to Theorem 2.14 {T, Y, Z} is the unique
Markov boundary of X.
Theorem 2.14 holds for all probability distributions including ones that are
not strictly positive. When a probability distribution is not strictly positive,
there is not necessarily a unique Markov boundary. This is shown in the following example:
Example 2.21 Let P be the probability distribution in Example 2.17. Then
{X} and {Y } are both Markov boundaries of {Z}. Note that neither DAG in
Figure 2.25 is faithful to P .
Our final result is that in the case of strictly positive distributions the Markov
boundary is unique:
Theorem 2.15 Suppose P is a strictly positive probability distribution of the
variables in V. Then for each X ∈ V there is a unique Markov boundary of X.
Proof. The proof can be found in [Pearl, 1988].
110
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
F
D
G
Figure 2.27: This DAG is not a minimal description of the probability distribution of the variables if the only influence of F on G is through D.
2.6
More on Causal DAGs
Recall from Section 1.4 that if we create a causal DAG G = (V, E) and assume
the probability distribution of the variables in V satisfies the Markov condition
with G, we say we are making the causal Markov assumption. In that section
we argued that, if we define causation based on manipulation, this assumption
is often justified. Next we discuss three related causal assumptions, namely
the causal minimality assumption, the causal faithfulness assumption, and the
causal embedded faithfulness assumption.
2.6.1
The Causal Minimality Assumption
If we create a causal DAG G = (V, E) and assume the probability distribution of
the variables in V satisfies the minimality condition with G, we say we are making the causal minimality assumption. Recall if P satisfies the minimality
condition with G, then P satisfies the Markov condition with G. So the causal
minimality assumption subsumes the causal Markov assumption. If we define
causation based on manipulation and we feel the causal Markov assumption is
justified, would we also expect this assumption to be justified? In general, it
seems we would. The only apparent exception to minimality could be if we
included an edge from X to Y when X is only an indirect cause of Y through
some other variable(s) in V. Consider again the situation concerning finasteride,
DHT level, and hair growth discussed in Section 1.4. We noted that DHT level
is a causal mediary between finasteride and hair growth with finasteride having
no other causal path to hair growth. We concluded that hair growth (G) is
independent of finasteride (F ) conditional on DHT level (D). Therefore, if we
represent the causal relationships among the variables by the DAG in Figure
2.27, the DAG would not be a minimal description of the probability distribution because we can remove the edge F → G and the Markov condition will still
be satisfied. However, since we’ve defined a causal DAG (See the beginning of
Section 1.4.2.) to be one that contains only direct causal influences, the DAG
containing the edge F → G is not a causal DAG according to our definition.
So, given our definition of a causal DAG, this situation is not really an exception
to the causal minimality assumption.
2.6. MORE ON CAUSAL DAGS
F
111
D
G
Figure 2.28: If D does not transmit an influence from F to G, this causal DAG
will not be faithful to the probability distribution of the variables.
2.6.2
The Causal Faithfulness Assumption
If we create a causal DAG G = (V, E) and assume the probability distribution
of the variables in V satisfies the faithfulness condition with G, we say we are
making the causal faithfulness assumption. Recall if P satisfies the faithfulness condition with G, then P satisfies the minimality condition with G. So
the causal faithfulness assumption subsumes the causal minimality assumption.
If we define causation based on manipulation and we feel the causal minimality
assumption is justified, would we also expect this assumption to be justified? It
seems in most cases we would. For example, if the manipulation of X leads to
a change in the probability distribution of Y and to a change in the probability
distribution of Z, we would ordinarily not expect Y and Z to be independent.
That is, we ordinarily expect the presence of one effect of a cause should make
it more likely its other effects are present. Similarly, if the manipulation of X
leads to a change in the probability distribution of Y , and the manipulation of
Y leads to a change in the probability distribution of Z, we would ordinarily not
expect X and Z to be independent. That is, we ordinarily expect a causal mediary to transmit an influence from its antecedent to its consequence. However,
there are notable exceptions. Recall in Section 1.4.1 we offered the possibility
that a certain minimal level of DHT is necessary for hair loss, more than that
minimal level has no further effect on hair loss, and finasteride is not capable of
lowering DHT level below that level. That is, it may be that finasteride (F ) has
a causal effect on DHT level (D), DHT level has a causal effect on hair growth
(G), and yet finasteride has no effect on hair growth. Our causal DAG, which
is shown in Figure 2.28, would then not be faithful to the distribution of the
variables because its structure does not entail IP ({G}, {F }). Figure 2.20 shows
actual probability values which result in this independence. Recall that it is not
even possible to faithfully embed the distribution, which is the product of the
conditional distributions shown in that figure.
This situation is fundamentally different than the problem encountered when
we fail to identify a hidden common cause (discussed in Section 1.4.2 and more
in the following subsection). If we fail to identify a hidden common cause,
our problem is in our lack of identifying variables; and, if we did successfully
identify all hidden common causes, we would ordinarily expect the Markov
condition, and indeed the faithfulness condition, to be satisfied. In the current
situation, the lack of faithfulness is inherent in the relationships among the
variables themselves. There are other similar notable exceptions to faithfulness.
Some are discussed in the exercises.
112
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
X
Y
X
Z
W
Z
H
Y
X
Y
W
Z
W
S
S
S
(a)
(b)
(c)
Figure 2.29: We would not expect the DAG in (a) to satisfy the Markov condition with the probability distribution of the 5 variables in that figure if Z and
W had a hidden cause, as depicted by the shaded node H in (b). We would
expect the DAG in (c) to be a minimal description of the distribution but not
faithful to it.
2.6.3
The Causal Embedded Faithfulness Assumption
In Section 1.4.2, we noted three important exceptions to the causal Markov assumptions. The first is that their can be no hidden common causes; the second
is that selection bias cannot be present; and the third is that there can be no
causal feedback loops. Since the causal faithfulness assumption subsumes the
causal Markov assumption, these are also exceptions to the causal faithfulness
assumption. As discussed in the previous subsection, other exceptions to the
causal faithfulness assumption include situations such as when a causal mediary fails to transmit an influence from its antecedent to its consequence. Of
these exceptions, the first exception (hidden common causes) seems to be most
prominent. Let’s discuss that exception further.
Suppose we identify the following causal relationships with manipulation:
X causes Z
Y causes W
Z causes S
W causes S.
Then we would construct the causal DAG shown in Figure 2.29 (a). The Markov
condition entails IP (Z, W ) for that DAG. However, if Z and W had a hidden
common cause as shown in Figure 2.29 (b), we would not ordinarily expect this
independency. This was discussed in Section 1.4.2. So if we fail to identify
a hidden common cause, ordinarily we would not expect the causal DAG to
satisfy the Markov condition with the probability distribution of the variables,
EXERCISES
113
which means it would not satisfy the faithfulness condition with that distribution either. However, we would ordinarily expect faithfulness to the DAG
that included all hidden common causes. For example, if H is the only hidden
common cause among the variables in the DAG in Figure 2.29 (b), we would
ordinarily expect the probability distribution of all six variables to satisfy the
faithfulness condition with the DAG in that figure, which means the probability
distribution of X, Y , Z, W , and S is embedded faithfully in that DAG. If we
assume the probability distribution of the observed variables is embedded faithfully in a causal DAG containing these variables and all hidden common causes,
we say we are making the causal embedded faithfulness assumption. It
seems this assumption is often justified. Perhaps the most notable exception
to it is the presence of selection bias. This exception is discussed further in
Exercise 2.35 and in Section 9.1.2.
Note that if we assume faithfulness to the DAG in Figure 2.29 (b), and we
add the adjacencies Z → W and X → W to the DAG in Figure 2.29 (a), the
probability distribution of S, X, Y , Z, and W would satisfy the Markov condition with the resultant DAG (shown in Figure 2.29 (c)) because this new DAG
does not entail IP ({Z}, {W }) or any other independencies not entailed by the
DAG in Figure 2.29 (b). The problem with the DAG in Figure 2.29 (c) is that it
fails to entail independencies that are present. That is, we have IP ({X}, {W }),
and the DAG in Figure 2.29 (c) does not entail this independency (Can you find
others that it fails to entail?). This means it is not faithful to the probability
distribution of S, X, Y , Z, and W . Indeed, similar to the result obtained in
Example 2.11, no DAG is faithful to the distribution of only S, X, Y , Z, and
W . Rather this distribution can only be embedded faithfully as done in Figure
2.29 (b) with the hidden common cause. Regardless, the DAG in Figure 2.29
(c) is a minimal description of the distribution of only S, X, Y , Z, and W , and
it constitutes a Bayesian network with that distribution. So any inference algorithms for Bayesian networks (discussed in Chapters 3, 4 and 5) are applicable
to it. However, it is no longer a causal DAG.
EXERCISES
Section 2.1
Exercise 2.1 Consider the DAG G in Figure 2.2. Prove that the Markov condition entails IP ({C}, {G}|{A, F }) for G.
Exercise 2.2 Suppose we add another variable R, an edge from F to R, and
an edge from R to C to the DAG G in Figure 2.3. The variable R might represent the professor’s initial reputation. State which of the following conditional
independencies you would feel are entailed by the Markov condition for G. For
each that you feel is entailed, try to prove it actually is.
114
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
1. Ip ({R}, {A}).
2. IP ({R}, {A}|{F }).
3. IP ({R}, {A}|{F, C}).
Exercise 2.3 State which of the following d-separations are in the DAG in
Figure 2.5:
1. IG ({W }, {S}|{Y, X}).
2. IG ({W }, {S}|{Y, Z}).
3. IG ({W }, {S}|{R, X}).
4. IG ({W, X}, {S, T }|{R, Z}).
5. IG ({Y, Z}, {T }|{R, S}).
6. IG ({X, S}, {W, T }|{R, Z}).
7. IG ({X, S, Z}, {W, T }|{R}).
8. IG ({X, Z}, {W }).
9. IG ({X, S, Z}, {W }).
Are {X, S, Z} and {W } d-separated by any set in that DAG?
Exercise 2.4 Let A, B, and C be subsets of a set of random variables V. Show
the following:
1. If A ∩ B = ∅, A ∩ C 6= ∅, and B ∩ C 6= ∅, then IP (A, B|C) is equivalent
to IP (A − C, B − C|C). That is, for every probability distribution P of V,
IP 0 (A, B|C) holds if and only IP (A − C, B − C|C) holds.
2. If A ∩ B 6= ∅ and P is a probability distribution of V such that IP (A, B|C)
holds, P is not positive definite. A probability distribution is positive
definite if there are no 0 values in the distribution.
3. If the Markov condition entails a conditional independency, then the independency must hold in a positive definite distribution. Hint: Use Theorem
1.5.
Conclude Lemma 2.2 from these three facts.
Exercise 2.5 Show IP ({X}, {Z}) for the distribution P in the Bayesian network in Figure 2.6.
Exercise 2.6 Use Algorithm 2.1 to find all nodes reachable from M in Figure
2.7. Show the labeling of the edges according to that algorithm.
EXERCISES
115
Exercise 2.7 Implement Algorithm 2.1 in the computer language of your choice.
Exercise 2.8 Perform a more rigorous analysis of Algorithm 2.1 than that done
in the text. That is, first identify basic operations. Then show W (m, n) ∈
O(mn) for these basic operations, and develop an instance showing W (m, n) ∈
Ω(mn).
Exercise 2.9 Implement Algorithm 2.2 in the computer language of your choice.
Exercise 2.10 Construct again a DAG representing the causal relationships
described in Exercise 1.25, but this time include auxiliary parent variables representing the possible values of the parameters in the conditional distributions.
Suppose we use the following variable names:
A:
B:
D:
L:
H:
T:
C:
Visit to Asia
Bronchitis
Dyspnea
Lung Cancer
Smoking History
Tuberculosis.
Chest X-ray
Identify the auxiliary parent variables, whose values we need to ascertain, for
each of the following calculations:
1. P ({B}|{H, D}).
2. P ({L}|{H, D}).
3. P ({T }|{H, D}).
Section 2.2
Exercise 2.11 Prove Corollary 2.1.
Exercise 2.12 In Part 2 of Case 1 in the proof of Theorem 2.4 it was left as
an exercise to show that if there is also a head-to-head meeting on σ2 at N 00 ,
there must be a head-to-head meeting at a node N ∗ ∈ N somewhere between N 0
and N 00 (including N 00 ) on ρ2 , and there cannot be a head-to-head meeting at
N ∗ on ρ1 . Show this. Hint: Recall ρ1 is not blocked.
Exercise 2.13 Show all DAGs Markov equivalent to each of the following DAGs,
and show the pattern representing the Markov equivalence class to which each
of the following belongs:
1. The DAG in Figure 2.15 (a).
2. The DAG in Figure 2.15 (c).
116
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
P(y1|x1) = b
P(y1|x2) = c
Y
P(x1) = a
X
W
P(w1|y1,z1) = d
P(w1|y1,z2) = e
P(w1|y2,z1) = e
P(w1|y2,z2) = f
Z
P(z1|x1) = c
P(z1|x2) = b
Figure 2.30: The probability distribution is not faithful to the DAG because
IP (W, X) and not IG (W, X). Each variable only has two possible values. So for
simplicity only the probability of one is shown.
3. The DAG in Figure 2.15 (d).
Exercise 2.14 Write a polynomial-time algorithm for determining whether two
DAGs are Markov equivalent. Implement the algorithm in the computer language of your choice.
Section 2.3
Exercise 2.15 Show that all the non-independencies listed in Example 2.8 hold
for the distribution discussed in that example.
Exercise 2.16 Assign arbitrary values to the conditional distributions for the
DAG in Figure 2.20, and see if the resultant distribution is faithful to the DAG.
Try to find an unfaithful distribution besides ones in the family shown in that
figure.
Exercise 2.17 Consider the Bayesian network in Figure 2.30.
1. Show that the probability distribution is not faithful to the DAG because
we have IP ({W }, {X}) and not IG ({W }, {X}).
2. Show further that this distribution does not admit a faithful DAG representation.
Exercise 2.18 Consider the Bayesian network in Figure 2.31.
EXERCISES
117
P(x1) = a
P(x2) = 1 - a
P(y1) = b
P(y2) = 1 - b
X
Y
Z
P(z1|x1,y1) = c
P(z2|x1,y1) = e
P(z3|x1,y1) = g
P(z4|x1,y1) = 1 - (c + e + g)
P(z1|x1,y2) = c
P(z2|x1,y2) = f
P(z3|x1,y2) = g
P(z4|x1,y2) = 1 - (c + f + g)
P(z1|x2,y1) = d
P(z2|x2,y1) = e
P(z3|x2,y1) = c + g - d
P(z4|x2,y1) = 1 - (c + e + g)
P(z1|x2,y2) = d
P(z2|x2,y2) = f
P(z3|x2,y2) = c + g - d
P(z4|x2,y2) = 1 - (c + f + g)
Figure 2.31: The probability distribution is not faithful to the DAG because
IP (X, Y |Z) and not IG (X, Y |Z).
1. Show that the probability distribution is not faithful to the DAG because
we have IP ({X}, {Y }|{Z}) and not IG ({X}, {Y }|{Z}).
2. Show further that this distribution does not admit a faithful DAG representation.
Exercise 2.19 Let V = {X, Y, Z, W ) and P be given by
P (x, y, z, w) = k × f(x, y) × g(y, z) × h(z, w) × i(w, x),
where f , g, h, and i are real-valued functions and k is a normalizing constant.
Show that this distribution does not admit a faithful DAG representation. Hint:
First show that the only conditional independencies are IP ({X}, {Z}|{Y, W })
and IP ({Y }, {W }|{X, Z}).
Exercise 2.20 Suppose we use the principle of indifference to assign probabilities to the objects in Figure 2.32. Let random variables V, S, C, L, and F be
defined as follows:
118
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
1
2
1
2
1
2
1
2
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
1
2
1
2
1
1
2
2
Figure 2.32: Objects with 5 properties.
Variable
V
S
C
L
F
Value
v1
v2
s1
s2
c1
c2
l1
l2
f1
f2
Outcomes Mapped to this Value
All objects containing a ‘1’
All objects containing a ‘2’
All square objects
All circular objects
All grey objects
All white objects
All objects covered with lines
All objects not covered with lines
All objects containing a number in a large font
All objects containing a number in a small font
Show that the probability distribution of V, S, C, L, and F is faithful to the DAG
in Figure 2.21 (a). The result in Example 2.11 therefore implies the marginal
distribution of V, S, L, and F is not faithful to any DAG.
Exercise 2.21 Prove Theorem 2.9.
Exercise 2.22 Prove Theorem 2.10.
Exercise 2.23 Develop a distribution, other than the one given in Example
2.11, which admits an embedded faithful DAG representation but does not admit
a faithful DAG representation.
Exercise 2.24 Show that the distribution discussed in Exercise 2.17 does not
admit an embedded faithful DAG representation.
Exercise 2.25 Show that the distribution discussed in Exercise 2.18 does not
admit an embedded faithful DAG representation.
EXERCISES
119
Exercise 2.26 Show that the distribution discussed in Exercise 2.19 does not
admit an embedded faithful DAG representation.
Section 2.4
Exercise 2.27 Obtain DAGs satisfying the minimality condition with P using
other orderings of the variables discussed in Example 2.16.
Section 2.5
Exercise 2.28 Apply Theorem 2.13 to find a Markov blanket for each node in
the DAG in Figure 2.26.
Exercise 2.29 Show that neither DAG in Figure 2.25 is faithful to the distribution discussed in Examples 2.17 and 2.21.
Section 2.6
Exercise 2.30 Besides IP ({X}, {W }), are there other independencies entailed
by the DAG in Figure 2.29 (b) that are not entailed by the DAG in Figure 2.29
(c)?
Exercise 2.31 Given the joint distribution of X, Y , Z, W , S, and H is faithful
to the DAG in Figure 2.29 (b), show that the marginal distribution of X, Y , Z,
W , and S does not admit a faithful DAG representation.
Exercise 2.32 Typing experience increases with age but manual dexterity decreases with age. Experience results in better typing performance as does good
manual dexterity. So it seems after an initial learning period, typing performance will stay about constant as age increases because the effects of increased
experience and decreased manual dexterity will cancel each other out. Draw
a DAG representing the causal influences among the variables, and discuss
whether the probability distribution of the variables is faithful to the DAG. If
it is not, show numeric values that could have this unfaithfulness. Hint: See
Exercise 2.17.
Exercise 2.33 Exercise 2.18 showed that the probability distribution in Figure
2.31 is not faithful to the DAG in that figure because IP ({X}, {Y }|{Z}) and
not IG ({X}, {Y }|{Z}). This means, if these are causal relationships, there is
no discounting (Recall discounting means one cause explains away a common
effect, thereby making the other cause less likely). Give an intuitive explanation
for why this might be the case. Hint: Note that the probability of each of Z’s
values is dependent on only one of the variables. For example, p(z1|x1, y1) =
p(z1|x1, y2) = p(z1|x1) and p(z1|x2, y1) = p(z1|x2, y2) = p(z1|x2).
120
CHAPTER 2. MORE DAG/PROBABILITY RELATIONSHIPS
Y
Z
X
W
S
Figure 2.33: Selection bias is present.
Exercise 2.34 The probability distribution in Figure 2.20 does not satisfy the
faithfulness condition with the DAG X ← Y → Z. Explain why. If these edges
describe causal influences, we would have two variables with a common cause
that are independent. Give an example for how this might happen.
Exercise 2.35 Suppose the probability distribution P of X, Y , Z, W , and S
is faithful to the DAG in Figure 2.33 and we are observing a subpopulation of
individuals who have S instantiated to a particular value s (as indicated by the
cross through S in the DAG). That is, selection bias is present (See Section
1.4.1.). Let P |s denote the probability distribution of X, Y , Z, and W conditional on S = s. Show that P |s does not admit an embedded faithful DAG
representation. Hint: First show that the only conditional independencies are
IP |s ({X}, {Z}|{Y, W }) and IP |s ({Y }, {W }|{X, Z}). Note that these are the
same conditional independencies as those obtained a different way in Exercise
2.19.
Part II
Inference
121
Chapter 3
Inference: Discrete
Variables
A standard application of Bayes’ Theorem (reviewed in Section 1.2) is inference
in a two-node Bayesian network. As discussed in Section 1.3, larger Bayesian
networks address the problem of representing the joint probability distribution of
a large number of variables and doing Bayesian inference with these variables.
For example, recall the Bayesian network discussed in Example 1.32. That
network, which is shown again in Figure 3.1, represents the joint probability
distribution of smoking history (H), bronchitis (B), lung cancer (L), fatigue
(F ), and chest X-ray (C).
If a patient had a smoking history and a positive chest X-ray, we would be
interested in the probability of that patient having lung cancer (i.e. P (l1|h1, c1))
and having bronchitis (i.e. P (b1|h1, c1)). In this chapter, we develop algorithms
that perform this type of inference.
In Section 3.1, we present simple examples showing why the conditional
independencies entailed by the Markov condition enable us to do inference with
a large number of variables. Section 3.2 develops Pearl’s [1986] message-passing
algorithm for doing exact inference in Bayesian networks. This algorithm passes
massages in the DAG to perform inference. In Section 3.3, we provide a version
of the algorithm that more efficiently handles networks in which the noisy orgate model is assumed. Section 3.4 references other inference algorithms that
also employ the DAG, while Section 3.5 presents the symbolic probabilistic
inference algorithm which does not employ the DAG. Next Section 3.6 discusses
the complexity of doing inference in Bayesian networks. Finally, Section 3.7
presents research relating Pearl’s message-passing algorithm to human causal
reasoning.
123
124
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
P(h1) = .2
H
P(b1|h1) = .25
P(b1|h2) = .05
B
L
F
P(f1|b1,l1) = .75
P(f1|b1,l2) = .10
P(f1|b2,l1) = .5
P(f1|b2,l2) = .05
P(l1|h1) = .003
P(l1|h2) = .00005
C
P(c1|l1) = .6
P(c1|l2) = .02
Figure 3.1: A Bayesian neworks. Each variable only has two values; so only the
probability of one is shown.
3.1
Examples of Inference
Next we present some examples illustrating how the conditional independencies
entailed by the Markov condition can be exploited to accomplish inference in a
Bayesian network.
Example 3.1 Consider the Bayesian network in Figure 3.2 (a). The prior
probabilities of all variables can be computed as follows:
P (y1) = P (y1|x1)P (x1) + P (y1|x2)P (x2) = (.9)(.4) + (.8)(.6) = .84
P (z1) = P (z1|y1)P (y1) + P (z1|y2)P (y2) = (.7)(.84) + (.4)(.16) = .652
P (w1) = P (w1|z1)P (z1) + P (w1|z2)P (z2) = (.5)(.652) + (.6)(.348) = .5348.
These probabilities are shown in Figure 3.2 (b). Note that the computation for
each variable requires information determined for its parent. We can therefore
consider this method a message passing algorithm in which each node passes
its child a message needed to compute the child’s probabilities. Clearly, this
algorithm applies to an arbitrarily long linked list and to trees.
Suppose next that X is instantiated for x1. Since the Markov condition
entails each variable is conditionally independent of X given its parent, we can
compute the conditional probabilities of the remaining variables by again passing
3.1. EXAMPLES OF INFERENCE
125
X
P(x1) = .4
X
P(x1) = .4
P(x2) = .6
Y
P(y1|x1) = .9
P(y1|x2) = .8
Y
P(y1) = .84
P(y2) = .16
Z
P(z1|y1) = .7
P(z1|y2) = .4
Z
P(z1) = .652
P(z2) = .348
W
P(w1|z1) = .5
P(w1|z2) = .6
W
P(w1) = .5348
P(w2) = .4652
(a)
(b)
Figure 3.2: A Bayesian network is in (a), and the prior probabilities of the
variables in that network are in (b). Each variable only has two values; so only
the probability of one is shown in (a).
messages down as follows:
P (y1|x1) = .9
P (z1|x1) = P (z1|y1, x1)P (y1|x1) + P (z1|y2, x1)P (y2|x1)
= P (z1|y1)P (y1|x1) + P (z1|y2)P (y2|x1)
= (.7)(.9) + (.4)(.1) = .67
P (w1|x1) = P (w1|z1, x1)P (z1|x1) + P (w1|z2, x1)P (z2|x1)
= P (w1|z1)P (z1|x1) + P (w1|z2)P (z2|x1)
= P ((.8)(.67) + (.6)(.33) = .734.
Clearly, this algorithm also applies to an arbitrarily long linked list and to trees.
The preceding instantiation shows how we can use downward propagation
of messages to compute the conditional probabilities of variables below the instantiated variable. Suppose now that W is instantiated for w1 (and no other
variable is instantiated). We can use upward propagation of messages to compute the conditional probabilities of the remaining variables as follows. First we
126
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
use Bayes’ theorem to compute P (z1|w1):
P (z1|w1) =
(.5)(.652)
P (w1|z1)P (z1)
=
= .6096.
P (w1)
.5348
Then to compute P (y1|w1), we again apply Bayes’ Theorem as follows:
P (y1|w1) =
P (w1|y1)P (y1)
.
P (w1)
We cannot yet complete this computation because we do not know P (w1|y1).
However, we can obtain this value in the manner shown when we discussed
downward propagation. That is,
P (w1|y1) = (P (w1|z1)P (z1|y1) + P (w1|z2)P (z2|y1).
After doing this computation, also computing P (w1|y2) (because X will need this
latter value), and then determining P (y1|w1), we pass P (w1|y1) and P (w1|y2)
to X. We then compute P (w1|x1) and P (x1|w1) in sequence as follows:
P (w1|x1) = (P (w1|y1)P (y1|x1) + P (w1|y2)P (y2|x1)
P (x1|w1) =
P (w1|x1)P (x1)
.
P (w1)
It is left as an exercise to perform these computations. Clearly, this upward
propagation scheme applies to an arbitrarily long linked list.
The next example shows how to turn corners in a tree.
Example 3.2 Consider the Bayesian network in Figure 3.3. Suppose W is instantiated for w1. We compute P (y1|w1) followed by P (x1|w1) using the upward
propagation algorithm just described. Then we proceed to compute P (z1|w1) followed by P (t1|w1) using the downward propagation algorithm. It is left as an
exercise to do this.
3.2
Pearl’s Message-Passing Algorithm
By exploiting local independencies as we did in the previous subsection, Pearl
[1986, 1988] developed a message-passing algorithm for inference in Bayesian
networks. Given a set a of values of a set A of instantiated variables, the algorithm determines P (x|a) for all values x of each variable X the network. It
accomplishes this by initiating messages from each instantiated variable to its
neighbors. These neighbors in turn pass messages to their neighbors. The updating does not depend on the order in which we initiate these messages, which
means the evidence can arrive in any order. First we develop the algorithm for
Bayesian networks whose DAGs are rooted trees; then we extend the algorithm
to singly-connected networks.
3.2. PEARL’S MESSAGE-PASSING ALGORITHM
127
P(x1) = .1
X
P(y1|x1) = .6
P(y1|x2) = .2
Y
Z
W
P(z1|x1) = .7
P(z1|x2) = .1
T
P(w1|y1) = .9
P(w1|y2) = .3
P(t1|z1) = .8
P(t1|z2) = .1
Figure 3.3: A Bayesian network that is a tree. Each variable only has two
possible values. So only the probability of one is shown.
3.2.1
Inference in Trees
Recall a rooted tree is a DAG in which there is a unique node called the root,
which has no parent, every other node has precisely one parent, and every node
is a descendent of the root.
The algorithm is based on the following theorem. It may be best to read the
proof of the theorem before its statement as its statement is not very transparent
without seeing it developed.
Theorem 3.1 Let (G, P ) be a Bayesian network whose DAG is a tree, where
G = (V, E), and a be a set of values of a subset A ⊂ V. For each variable X,
define λ messages, λ values, π messages, and π values as follows:
1. λ messages:
For each child Y of X, for all values of x,
X
P (y|x)λ(y).
λY (x) ≡
y
2. λ values:
If X ∈ A and X’s value is x̂,
λ(x̂) ≡ 1
λ(x) ≡ 0
for x 6= x̂.
If X ∈
/ A and X is a leaf, for all values of x,
λ(x) ≡ 1.
128
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
If X ∈
/ A and X is a nonleaf, for all values of x,
Y
λU (x),
λ(x) ≡
U ∈CHX
where CHX denotes the set of children of X.
3. π messages:
If Z is the parent of X, then for all values of z,
Y
λU (z).
πX (z) ≡ π(z)
U∈CHZ −{X}
4. π values:
If X ∈ A and X’s value is x̂,
π(x̂) ≡ 1
π(x) ≡ 0
for x 6= x̂.
If X ∈
/ A and X is the root, for all values of x,
π(x) ≡ P (x).
If X ∈
/ A, X is not the root, and Z is the parent of X, for all values of
x,
X
P (x|z)π X (z).
π(x) ≡
z
5. Given the definitions above, for each variable X, we have for all values of
x,
P (x|a) = αλ(x)π(x),
where α is a normalizing constant.
Proof. We will prove the theorem for the case where each node has precisely
two children. The case of an arbitrary tree is then a straightforward generalization. Let DX be the subset of A containing all members of A that are in the
subtree rooted at X (therefore, including X if X ∈ A), and NX be the subset
of A containing all members of A that are nondescendents of X. Recall X is a
nondescendent of X; so this set includes X if X ∈ A. This situation is depicted
in Figure 3.4. We have for each value of x,
P (x|a) = P (x|dX , nX )
P (dX , nX |x)P (x)
=
P (dX , nX )
P (dX |x)P (nX |x)P (x)
=
P (dX , nX )
P (dX |x)P (x|nX )P (nX )P (x)
=
P (x)P (dX , nX )
= βP (dX |x)P (x|nX ),
(3.1)
3.2. PEARL’S MESSAGE-PASSING ALGORITHM
129
NX
X
DX
Figure 3.4: The set of instantiated variables A = NX ∪ DX . If X ∈ A, X is in
both NX and DX .
where β is a constant that does not depend on the value of x. The 2nd and
4th equalities are due to Bayes’ Theorem. The 3rd equality follows directly from
d-separation (Lemma 2.1) if X ∈
/ A. It is left as an exercise to show it still
holds if X ∈ A.
We will develop functions λ(x) and π(x) such
λ(x) w P (dX |x)
π(x) w P (x|nX ).
By w we mean ‘proportional to’. That is, π(x), for example, may not equal
P (x|nX ), but it equals a constant times P (x|nX ), where the constant does not
depend on the value of x. Once we do this, due to Equality 3.1, we will have
P (x|a) = αλ(x)π(x),
where α is a normalizing constant that does not depend on the value of x.
1. Develop λ(x): We need
λ(x) w P (dX |x).
(3.2)
Case 1: X ∈ A and X’s value is x̂. Since X ∈ DX ,
P (dX |x) = 0
for x 6= x̂.
So to achieve Proportionality 3.2, we can set
λ(x̂) ≡ 1
λ(x) ≡ 0
for x 6= x̂.
Case 2: X ∈
/ A and X is a leaf. In this case dX = ∅, the empty set of
variables, and so
P (dX |x) = P (∅|x) = 1
for all values of x.
130
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
X
DX
Y
W
DY
DW
Figure 3.5: If X is not in A, then DX = DY ∪ DW .
So to achieve Proportionality 3.2, we can set
λ(x) ≡ 1
for all values of x.
Case 3: X ∈
/ A and X is a nonleaf. Let Y be X’s left child, W be X’s
right child. Then since X ∈
/ A,
DX = DY ∪ DW .
This situation is depicted in Figure 3.5. We have
P (dX |x) = P (dY , dW |x)
= P (dY |x)P (dW |x)
X
X
P (y|x)P (dY |y)
P (w|x)P (dW |w)
=
y
w
X
P (y|x)λ(y)
y
X
w
P (w|x)λ(w).
w
The second equality is due to d-separation and the third to the law of
total probability. So we can achieve Proportionality 3.2 by defining
for all values of x,
X
P (y|x)λ(y)
λY (x) ≡
y
λW (x) ≡
X
P (w|x)λ(w),
w
and setting
λ(x) ≡ λY (x)λW (x)
for all values of x.
3.2. PEARL’S MESSAGE-PASSING ALGORITHM
131
NZ
Z
T
NX
X
DT
Figure 3.6: If X is not in E, then NX = NZ ∪ DT .
2. Develop π(x): We need
π(x) w P (x|nX ).
(3.3)
Case 1: X ∈ A and X’s value is x̂. Due to the fact that X ∈ NX ,
P (x̂|nX ) = P (x̂|x̂) = 1
P (x|nX ) = P (x|x̂) = 0
for x 6= x̂.
So we can achieve Proportionality 3.3 by setting
π(x̂) ≡ 1
π(x) ≡ 0
for x 6= x̂.
Case 2: X ∈
/ A and X is the root. In this case nX = ∅, the empty set
of random variables, and so
P (x|nX ) = P (x|∅) = P (x)
for all values of x.
So we can achieve Proportionality 3.3 by setting
π(x) ≡ P (x)
for all value of x.
Case 3: X ∈
/ A and X is not the root. Without loss of generality assume
X is Z’s right child, and let T be Z’s left child. Then NX = NZ ∪DT .
132
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
This situation is depicted in Figure 3.6. We have
X
P (x|nX ) =
P (x|z)P (z|nX )
z
=
X
P (x|z)P (x|nZ , dT )
z
=
X
P (x|z)
z
= γ
X
P (z|nZ )P (nZ )P (dT |z)P (z)
P (z)P (nZ , dT )
P (x|z)π(z)λT (z).
z
It is left as an exercise to obtain the third equality above using the
same manipulations as in the derivation of Equality 3.1. So we can
achieve Proportionality 3.3 by defining for all values of z,
π X (z) ≡ π(z)λT (z),
and setting
π(x) ≡
X
P (x|z)πX (z)
for all values of x.
z
This completes the proof.
Next we present an algorithm based on this theorem. It is left as an exercise
to show its correctness follows from the theorem. Clearly, the algorithm can be
implemented as an object-oriented program, in which each node is an object that
communicates with the other nodes by passing λ and π messages. However, our
goal is to show the steps in the algorithm rather than to discuss implementation.
So we present it using top-down design.
Before presenting the algorithm, we show how the routines in it are called.
Routine initial_tree is first called as follows:
initial_tree((G, P ), A, a, P (x|a));
After this call, A and a are both empty, and for every variables X, for every
value of x, P (x|a) is the conditional probability of x given a, which, since a is
empty, is the prior probability of x. Each time a variable V is instantiated for
v̂, routine update-tree is called as follows:
update_tree((G, P ), A, a, V, v̂, P (x|a));
After this call, V has been added to A, v̂ has been added to a, and for every
variables X, for every value of x, P (x|a) has been updated to be the conditional
probability of x given the new value of a. The algorithm now follows.
3.2. PEARL’S MESSAGE-PASSING ALGORITHM
133
Algorithm 3.1 Inference-in-Trees
Problem: Given a Bayesian network whose DAG is a tree, determine the
probabilities of the values of each node conditional on specified values of
the nodes in some subset.
Inputs: Bayesian network (G, P ) whose DAG is a tree, where G = (V, E), and
a set of values a of a subset A ⊆ V.
Outputs: The Bayesian network (G, P ) updated according to the values in
a. The λ and π values and messages and P (x|a) for each X ∈ V are
considered part of the network.
void initial_tree (Bayesian-network& (G, P ) where G = (V, E),
set-of-variables& A, set-of-variable-values& a)
{
A = ∅; a = ∅;
for (each X ∈ V) {
for (each value x of X)
λ(x) = 1;
// Compute λ values.
for (the parent Z of X)
// Does nothing if X is the a root.
for (each value z of Z)
// Compute λ messages.
λX (z) = 1;
}
for (each value r of the root R) {
P (r|a) = P (r);
// Compute P (r|a).
π(r) = P (r);
// Compute R’s π values.
}
for (each child X of R)
send_π_msg(R, X);
}
void update_tree (Bayesian-network& (G, P ) where G = (V, E),
set-of-variables& A, set-of-variable-values& a,
variable V , variable-value v̂)
{
A = A ∪ {V }; a = a ∪ {v̂};
// Add V to A.
λ(v̂) = 1; π(v̂) = 1; P (v̂|a) = 1;
// Instantiate V to v̂.
for (each value of v 6= v̂) {
λ(v) = 0; π(v) = 0; P (v|a) = 0;
}
if (V is not the root && V ’s parent Z ∈
/ A)
send_λ_msg(V, Z);
for (each child X of V such that X ∈
/ A)
send_π_msg(V, X);
}
134
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
void send_λ_msg(node Y , node X)
{
for (each value
P of x) {
λY (x) = P (y|x)λ(y);
// For simplicity (G, P ) is
// not shown as input.
// Y sends X a λ message.
y
Q
λ(x) =
λU (x);
// Compute X’s λ values.
U ∈CHX
P (x|a) = αλ(x)π(x);
// Compute P (x|a).
}
normalize P (x|a);
if (X is not the root and X’s parent Z ∈
/ A)
send_λ_msg(X, Z);
for (each child W of X such that W 6= Y and W ∈
/ A)
send_π_msg(X, W );
}
void send_π_msg(node Z, node X)
{
for (each value of z) Q
λY (z);
πX (z) = π(z)
// For simplicity (G, P ) is
// not shown as input.
// Z sends X a π message.
Y ∈CHZ −{X}
for (each P
value of x) {
π(x) = P (x|z)πX (z);
// Compute X’s π values.
z
P (x|a) = αλ(x)π(x);
}
normalize P (x|a);
for (each child Y of X such that Y ∈
/ A)
send_π_msg(X, Y );
// Compute P (x|a).
}
Examples of applying the preceding algorithm follow:
Example 3.3 Consider the Bayesian network in Figure 3.7 (a). It is the network in Figure 3.1 with node F removed. We will show the steps when the
network is initialized.
The call
initial_tree((G, P ), A, a);
results in the following steps:
3.2. PEARL’S MESSAGE-PASSING ALGORITHM
135
P(h1) = .2
H
P(b1|h1) = .25
P(b1|h2) = .05
B
L
P(l1|h1) = .003
P(l1|h2) = .00005
C
P(c1|l1) = .6
P(c1|l2) = .02
(a)
8(h) = (1,1)
B(h) = (.2,.8)
P(h|i) = (.2,.8)
88B(h) = (1,1)
9BB(h) = (.2,.8)
8(b) = (1,1)
B(b) = (.09,.91)
P(b|i) = (.09,.91)
H
B
88L(h) = (1,1)
9BL(h) = (.2,.8)
L
8(l) = (1,1)
B(l) = (.00064,.99936)
P(l|i) = (.00064,.99936)
88C(l) = (1,1)
9BC(l) = (.00064,.99936)
C
8(c) = (1,1)
B(c) = (.02037,.97963)
P(c|i) = (.02037,.97963)
(b)
Figure 3.7: Figure (b) shows the initialized network corresponding to the
Bayesian network in Figure (a). In Figure (b) we write, for example, P (h|∅) =
(.2, .8) instead of P (h1|∅) = .2 and P (h2|∅) = .8.
136
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
A = ∅;
a = ∅;
λ(h1) = 1; λ(h2) = 1;
λ(b1) = 1; λ(b2) = 1;
λ(l1) = 1; λ(l2) = 1;
λ(c1) = 1; λ(c2) = 1;
// Compute λ values.
λB (h1) = 1; λB (h2) = 1;
λL (h1) = 1; λL (h2) = 1;
λC (l1) = 1; λC (l2) = 1;
// Compute λ messages.
P (h1|∅) = P (h1) = .2;
P (h2|∅) = P (h2) = .8;
// Compute P (h|∅).
π(h1) = P (h1) = .2;
π(h2) = P (h2) = .8;
// Compute H’s π values.
send_π_msg(H, B);
send_π_msg(H, L);
The call
send_π_msg(H, B);
results in the following steps:
πB (h1) = π(h1)λL (h1) = (.2)(1) = .2;
πB (h2) = π(h2)λL (h2) = (.8)(1) = .8;
// H sends B a π message.
π(b1) = P (b1|h1)π B (h1) + P (b1|h2)π B (h2);
= (.25)(.2) + (.05)(.8) = .09;
// Compute B’s π values.
π(b2) = P (b2|h1)π B (h1) + P (b2|h2)π B (h2);
= (.75)(.2) + (.95)(.8) = .91;
P (b1|∅) = αλ(b1)π(b1) = α(1)(.09) = .09α;
P (b2|∅) = αλ(b2)π(b2) = α(1)(.91) = .91α;
P (b1|∅) =
.09α
.09α+.91α
= .09;
P (b1|∅) =
.91α
.09α+.91α
= .91;
The call
send_π_msg(H, L);
// Compute P (b|∅).
3.2. PEARL’S MESSAGE-PASSING ALGORITHM
137
results in the following steps:
π L (h1) = π(h1)λB (h1) = (.2)(1) = .2;
π L (h2) = π(h2)λB (h2) = (.8)(1) = .8;
// H sends L a π
// message.
π(l1) = P (l1|h1)πL (h1) + P (l1|h2)πL (h2);
= (.003)(.2) + (.00005)(.8) = .00064;
// Compute L’s π
// values.
π(l2) = P (l2|h1)πL (h1) + P (l2|h2)πL (h2);
= (.997)(.2) + (.99995)(.8) = .99936;
P (l1|∅) = αλ(l1)π(l1) = α(1)(.00064) = .00064α;
P (l2|∅) = αλ(l2)π(l2) = α(1)(.99936) = .99936α;
P (l1|∅) =
.00064α
.00064α+.99936α
= .00064;
P (l1|∅) =
.99936α
.00064α+.99936α
= .99936;
// Compute P (l|∅).
send_π_msg(L, C);
The call
send_π_msg(L, C);
results in the following steps:
π C (l1) = π(l1) = .00064;
π C (l2) = π(l2) = .99936;
// L sends C a π.
// message.
π(c1) = P (c1|l1)πC (l1) + P (c1|l2)πC (l2);
= (.6)(.00064) + (.02)(.99936) = .02037;
// Compute C’s π
// values.
π(c2) = P (c2|l1)πC (l1) + P (c2|l2)πC (l2);
= (.4)(.00064) + (.98)(.99936) = .97963;
P (c1|∅) = αλ(c1)π(c1) = α(1)(.02037) = .02037α;
P (c2|∅) = αλ(c2)π(c2) = α(1)(.97963) = .97963α;
P (c1|∅) =
.02037α
.02037α+.97963α
= .02037;
P (c1|∅) =
.97963α
.02037α+.97963α
= .97963;
// Compute P (c|∅).
The initialization is now complete. The initialized network is shown in Figure
3.7 (b).
138
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
Example 3.4 Consider again the Bayesian network in Figure 3.7 (a). Suppose
B is instantiated for b1. That is, we find out the patient has bronchitis. Next we
show the steps in the algorithm when the network’s values are updated according
to this instantiation.
The call
update_tree((G, P ), A, a, B, b1);
results in the following steps:
A = ∅ ∪ {B} = {B};
a = ∅ ∪ {b1} = {b1};
λ(b1) = 1; π(b1) = 1; P (b1|{b1}) = 1;
λ(b2) = 0; π(b2) = 0; P (b2|{b1}) = 0;
// Instantiate B for b1.
send_λ_msg(B, H);
The call
send_λ_msg(B, H);
results in the following steps:
λB (h1) = P (b1|h1)λ(b1) + P (b2|h1)λ(b2);
= (.25)(1) + .75(0) = .25;
// B sends H a λ
// message.
λB (h2) = P (b1|h2)λ(b1) + P (b2|h2)λ(b2);
= (.05)(1) + .95(0) = .05;
λ(h1) = λB (h1)λL (h1) = (.25)(1) = .25;
λ(h2) = λB (h2)λL (h2) = (.05)(1) = .05;
// Compute H’s λ
// values.
P (h1|{b1}) = αλ(h1)π(h1) = α(.25)(.2) = .05α;
P (h2|{b1}) = αλ(h2)π(h2) = α(.05)(.8) = .04α;
// Compute P (h|{b1}).
P (h1|{b1}) =
.05α
.05α+.04α
= .5556;
P (h2|{b1}) =
.04α
.04α+.05α
= .4444;
send_π_msg(H, L);
The call
send_π_msg(H, L);
3.2. PEARL’S MESSAGE-PASSING ALGORITHM
139
results in the following steps:
π L (h1) = π(h1)λB (h1) = (.2)(.25) = .05;
π L (h2) = π(h2)λB (h2) = (.8)(.05) = .04;
// H sends L a π
// message.
π(l1) = P (l1|h1)πL (h1) + P (l1|h2)πL (h2);
= (.003)(.05) + (.00005)(.04) = .00015;
// Compute L’s π
// values.
π(l2) = P (l2|h1)πL (h1) + P (l2|h2)πL (h2);
= (.997)(.05) + (.99995)(.04) = .08985;
P (l1|{b1}) = αλ(l1)π(l1) = α(1)(.00015) = .00015α;
P (l2|{b1}) = αλ(l2)π(l2) = α(1)(.08985) = .08985α;
P (l1|{b1}) =
.00015α
.00015α+.08985α
= .00167;
P (l2|{b1}) =
.00015α
.00015α+.08985α
= .99833;
// Compute
// P (l|{b1}).
send_π_msg(L, C);
The call
send_π_msg(L, C);
results in the following steps:
π C (l1) = π(l1) = .00015;
π C (l2) = π(l2) = .08985;
// L sends C a π
// message.
π(c1) = P (c1|l1)πC (l1) + P (c1|l2)πC (l2);
= (.6)(.00015) + (.02)(.08985) = .00189;
// Compute C’s π
// values.
π(c2) = P (c2|l1)πC (l1) + P (c2|l2)πC (l2);
= (.4)(.00015) + (.98)(.08985) = .08811;
P (c1|{b1}) = αλ(c1)π(c1) = α(1)(.00189) = .00189α;
P (c2|{b1}) = αλ(c2)π(c2) = α(1)(.08811) = .08811α;
P (l1|{b1}) =
.00189α
.00189α+.08811α
= .021;
P (l2|{b1}) =
.08811α
.00189α+.08811α
= .979;
// Compute
// P (c|{b1}).
The updated network in shown in Figure 3.8 (a). Notice that the probability of
lung cancer increases slightly when we find out the patient has bronchitis. The
reason is that they have the common cause smoking history, and the presence of
140
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
8(h) = (.25,.05)
B(h) = (.2,.8)
P(h|{b1}) = (.5556,.4444)
88B(h) = (.25,.05)
9BB(h) = (.2,.8)
8(b) = (1,0)
B(b) = (1,0)
P(b|{b1}) = (1,0)
H
B
88L(h) = (1,1)
9BL(h) = (.05,.04)
L
8(l) = (1,1)
B(l) = (.00015,.08985)
P(l|{b1}) = (.00167,.99833)
88C(l) = (1,1)
9BC(l) = (.00015,.08985)
C
8(c) = (1,1)
B(c) = (.00189,.08811)
P(c|{b1}) = (.021,.979)
(a)
8(h) = (.00544,.00100)
B(h) = (.2,.8)
P(h|{b1,c1}) = (.57672,.42328)
88B(h) = (.25,.05)
9BB(h) = (.2,.8)
8(b) = (1,0)
B(b) = (1,0)
P(b|{b1,c1}) = (1,0)
H
B
88L(h) = (.02174,.02003)
9BL(h) = (.05,.04)
L
8(l) = (.6,.02)
B(l) = (.00015,.08985)
P(l||{b1,c1}) = (.04762,.95238)
88C(l) = (.6,.02)
9BC(l) = (.00015,.08985)
C
8(c) = (1,0)
B(c) = (1,0)
P(c|{b1,c1}) = (1,0)
(b)
Figure 3.8: Figure (a) shows the updated network after B is instantiated for b1.
Figure (b) shows the updated network after B is instantiated for b1 and C is
instantiated for c1.
3.2. PEARL’S MESSAGE-PASSING ALGORITHM
141
bronchitis raises the probability of this cause, which in turn raises the probability
of its other effect lung cancer.
Example 3.5 Consider again the Bayesian network in Figure 3.7 (a). Suppose
B has already been instantiated for b1, and C is now instantiated for c1. That is,
we find out the patient has a positive chest X-ray. Next we show the steps in the
algorithm when the network’s values are updated according to this instantiation.
The call
update_tree((G, P ), A, a, C, c1);
results in the following steps:
A = {B} ∪ {C} = {B, C};
a = {b1} ∪ {c1} = {b1, c1};
λ(c1) = 1; π(c1) = 1; P (c1|{b1, c1}) = 1;
λ(c2) = 0; π(c2) = 0; P (c2|{b1, c1}) = 0;
// Instantiate C for c1.
send_λ_msg(C, L);
The call
send_λ_msg(C, L);
results in the following steps:
λC (l1) = P (c1|l1)λ(c1) + P (c2|l1)λ(c2);
= (.6)(1) + (.4)(0) = .6;
// C sends L a λ message.
λC (l2) = P (c1|l2)λ(c1) + P (c2|l2)λ(c2);
= (.02)(1) + .98(0) = .02;
λ(l1) = λC (l1) = .6;
λ(l2) = λC (l2) = .02;
// Compute L’s λ values.
P (l1|{b1, c1}) = αλ(l1)π(l1) = α(.6)(.00015) = .00009α;
P (l2|{b1, c1}) = αλ(l2)π(l2) = α(.02)(.08985) = .00180α;
P (l1|{b1, c1}) =
.00009α
.00009α+.00180α
= .04762;
P (l2|{b1, c1}) =
.00180α
.00009α+.00180α
= .95238;
send_λ_msg(L, H);
The call
// Compute P (l|{b1, c1}).
142
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
send_λ_msg(L, H);
results in the following steps:
λL (h1) = P (l1|h1)λ(l1) + P (l2|h1)λ(l2);
= (.003)(.6) + .997(.02) = .02174;
// L sends H a λ
// message.
λL (h2) = P (l1|h2)λ(l1) + P (l2|h2)λ(l2);
= (.00005)(.6) + .99995(.02) = .02003;
λ(h1) = λB (h1)λL (h1) = (.25)(.02174) = .00544;
λ(h2) = λB (h2)λL (h2) = (.05)(.02003) = .00100;
// Compute H’s λ
// values.
P (h1|{b1, c1}) = αλ(h1)π(h1) = α(.00544)(.2) = .00109α;
P (h2|{b1, c1}) = αλ(h2)π(h2) = α(.00100)(.8) = .00080α;
P (h1|{b1, c1}) =
.00109α
.00109α+.00080α
= .57672;
P (h2|{b1, c1}) =
.0008α
.00109α+.00080α
= .42328;
// Compute P (h|{b1, c1}).
The updated network is shown in Figure 3.8 (b).
3.2.2
Inference in Singly-Connected Networks
A DAG is called singly-connected if there is at most one chain between any
two nodes. Otherwise, it is called multiply-connected. A Bayesian network is
called singly-connected if its DAG is singly-connected and is called multiplyconnected otherwise. For example, the DAG in Figure 3.1 is not singlyconnected because there are two chains between a number of nodes including,
for example, between B and L. The difference between a singly-connected DAG,
that is not a tree, and a tree is that in the latter a node can have more than
one parent. Figure 3.9 shows a singly-connected DAG that is not a tree. Next
we present an extension of the algorithm for trees to one for singly-connected
DAGs. Its correctness is due to the following theorem, whose proof is similar to
the proof of Theorem 3.1.
Theorem 3.2 Let (G, P ) be a Bayesian network that is singly-connected, where
G = (V, E), and a be a set of values of a subset A ⊂ V. For each variable X,
define λ messages, λ values, π messages, and π values as follows:
1. λ messages:
For each child Y of X, for all values of x,
!#
"
Ã
k
X
Y
X
πY (wi )
λ(y).
P (y|x, w1 , w2 , . . . wk )
λY (x) ≡
y
w1 ,w2 ,...wk
i=1
where W1 , W2 , . . . , Wk are the other parents of Y .
3.2. PEARL’S MESSAGE-PASSING ALGORITHM
Figure 3.9: A singly-connected network that is not a tree.
2. λ values:
If X ∈ A and X’s value is x̂,
λ(x̂) ≡ 1
λ(x) ≡ 0
for x 6= x̂.
If X ∈
/ A and X is a leaf, for all values of x,
λ(x) ≡ 1.
If X ∈
/ A and X is a nonleaf, for all values of x,
Y
λ(x) ≡
λU (x).
U ∈CHX
where CHX is the set of all children of X.
3. π messages:
Let Z be a parent of X. Then for all values of z,
Y
π X (z) ≡ π(z)
λU (z).
U ∈CHZ −{X}
4. π values:
143
144
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
If X ∈ A and X’s value is x̂,
π(x̂) ≡ 1
π(x) ≡ 0
for x 6= x̂.
If X ∈
/ A and X is a root, for all values of x,
π(x) ≡ P (x).
If X ∈
/ A, X is a nonroot, and Z1 , Z2 , ... Zj are the parents of X, for
all values of x,
π(x) =
X
z1 ,z2 ,...zj
Ã
P (x|z1 , z2 , . . . zj )
j
Y
!
πX (zi ) .
i=1
5. Given the definitions above, for each variable X, we have for all values of
x,
P (x|a) = αλ(x)π(x),
where α is a normalizing constant.
Proof. The proof is left as an exercise.
The algorithm based on the preceding theorem now follows.
Algorithm 3.2 Inference-in-Singly-Connected-Networks
Problem: Given a singly-connected Bayesian network, determine the probabilities of the values of each node conditional on specified values of the
nodes in some subset.
Inputs: Singly-connected Bayesian network (G, P ), where G = (V, E), and a
set of values a of a subset A ⊆ V.
Outputs: The Bayesian network (G, P ) updated according to the values in
a. The λ and π values and messages and P (x|a) for each X ∈ V are
considered part of the network.
void initial_net (Bayesian-network& (G, P ) where G = (V, E),
set-of-variables& A, set-of-variable-values& a)
3.2. PEARL’S MESSAGE-PASSING ALGORITHM
145
{
A = ∅; a = ∅;
for (each X ∈ V) {
for (each value x of X)
λ(x) = 1;
for (each parent Z of X)
for (each value z of Z)
λX (z) = 1;
for (each child Y of X)
for (each value x of X)
π Y (x) = 1;
}
for each root R {
for each value of r {
P (r|a) = P (r);
π(r) = P (r);
}
for (each child X of R)
send_π_msg(R, X);
}
// Compute λ values.
// Does nothing if X is the a root.
// Compute λ messages.
// Initialize π messages.
// Compute P (r|a).
// Compute R’s π values.
}
void update_tree (Bayesian-network (G, P ) where G = (V, E),
set-of-variables& A, set-of-variable-values& a,
variable V , variable-value v̂)
{
A = A ∪ {V }; a = a∪{v̂};
λ(v̂) = 1; π(v̂) = 1; P (v̂|a) = 1;
for (each value of v 6= v̂) {
λ(v) = 0; π(v) = 0; P (v|a) = 0;
}
for (each parent Z of V such that Z ∈
/ A)
send_λ_msg(V, Z);
for (each child X of V )
send_π_msg(V, X);
// Add V to A.
// Instantiate V for v̂.
}
void send_λ_msg(node Y , node X) // (G, P ) is not shown as input.
{
// Wi s are Y ’s other parents.
for each value of x {
// Y sends X a λ message.
"
¶#
µ
k
P
Q
P
πY (wi ) λ(y);
λY (x) ≡
P (y|x, w1 , w2 , . . . wk )
y
w1 ,w2 ,...wk
i=1
146
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
λ(x) =
Q
λU (x);
// Compute X’s λ values.
U ∈CHX
P (x|a) = αλ(x)π(x);
// Compute P (x|a).
}
normalize P (x|a);
for (each parent Z of X such that Z ∈
/ A)
send_λ_msg(X, Z);
for (each child W of X such that W 6= Y )
send_π_msg(X, W );
}
void send_π_message(node Z, node X)
{
for (each value of z) Q
λY (z);
πX (z) = π(z)
// (G, P ) is not shown as
// input.
// Z sends X a π message.
Y ∈CHZ −{X}
if (X ∈
/ A) {
for (each value of x) {
// the Zi s are X’s parents.
¶
µ
j
P
Q
π(x) =
π X (zi ) ;
P (x|z1 , z2 , . . . zj )
z1 ,z2 ,...zj
P (x|a) = αλ(x)π(x);
}
normalize P (x|a);
for (each child Y of X)
send_π_msg(X, Y );
}
if not (λ(x) = 1 for all values of x)
for (each parent W of X
such that W 6= Z and W ∈
/ A)
send_λ_msg(X, W );
i=1
// Compute X’s π values.
// Compute P (x|a).
//
//
//
//
Do not send λ messages to
X’s other parents if X and
all of X’s descendents are
uninstantiated.
}
Notice that the comment in routine send-π-message says ‘do not send λ
messages to X’s other parents if X and all of X’s descendents are uninstantiated.’ The reason is that, if X and all X’s descendents are uninstantiated, X
d-separates each of its parents from every other parent. Clearly, if X and all
X’s descendents are uninstantiated, then all X’s λ values are still equal to 1.
Examples of applying the preceding algorithm follow.
Example 3.6 Consider the Bayesian network in Figure 3.10 (a). For the sake
of concreteness, suppose the variables are the ones discussed in Example 1.37.
That is, they represent the following:
3.2. PEARL’S MESSAGE-PASSING ALGORITHM
147
8(b) = (1,1)
B(b) = (.005,.995)
P(b1) = .005
P(f1) = .03
B
F
P(b|
8(f) = (1,1)
B(f) = (.03,.97)
i) = (.005,.995)
F
88A(f) = (1,1)
9B A(f) = (.03,.97)
A
A
P(a1|b1,f1) = .992
P(a1|b2,f1) = .2
P(a1|b1,f2) = .99
P(a1|b2,f2) = .003
8(h) = (1,1)
B(h) = (.014,.986)
P(h|
(a)
8(f) = (.204,.008)
B(f) = (.03,.97)
P(f|{a1}) = (.429,.571)
F
88A(f) = (.204,.008)
9B A(f) = (.03,.97)
A
8(a) = (1,0)
B(a) = (1,0)
P(a|{a1}) = (1,0)
(c)
i) = (.014,.986)
(b)
B
88A (b) = (.990,.009)
9BA (b) = (.005,.995)
i) = (.03,.97)
B
88A(b) = (1,1)
9B A(b) = (.005,.995)
8(b) = (.990,.009)
B(b) = (.005,.995)
P(b|{a1}) = (.357,.643)
P(f|
8(b) = (.992,.2)
B(b) = (.005,.995)
P(b|{a1,f1}) = (.025,.975)
8(f) = (1,0)
B(f) = (1,0)
P(f|{a1}) = (1,0)
B
F
88A(b) = (.992,.2)
9B A(b) = (.005,.995)
88A(f) = (.204,.008)
9BA(f) = (1,0)
A
8(a) = (1,0)
B(a) = (1,0)
P(a|{a1}) = (1,0)
(d)
Figure 3.10: Figure (b) shows the initialized network corresponding to the
Bayesian network in Figure (a). Figure (c) shows the state of the network after
A is instantiated for a1, and Figure (d) shows its state after A is instantiated
for a1 and F is instantiated for f 1.
148
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
Variable
B
F
A
Value
b1
b2
f1
f2
a1
m2
When the Variable Takes this Value
A burglar breaks in house
A burglar does not break in house
Freight truck makes a delivery
Freight truck does not make a delivery
Alarm sounds
Alarm does not sound
We show the steps when the network is initialized.
The call
initial_tree((G, P ), A, a);
results in the following steps:
A = ∅;
a = ∅;
λ(b1) = 1; λ(b2) = 1;
λ(f 1) = 1; λ(f2) = 1;
λ(a1) = 1; λ(a2) = 1;
// Compute λ values.
λA (b1) = 1; λA (b2) = 1;
λA (f 1) = 1; λA (f2) = 1;
// Compute λ messages.
πA (b1) = 1; πA (b2) = 1;
πA (f 1) = 1; πA (f2) = 1;
// Compute π messages.
P (b1|∅) = P (b1) = .005;
P (b2|∅) = P (b2) = .995;
// Compute P (b|∅).
π(b1) = P (b1) = .005;
π(b2) = P (b2) = .995;
// Compute B’s π values.
send_π_msg(B, A);
P (f 1|∅) = P (f1) = .03;
P (f 2|∅) = P (f2) = .97;
// Compute P (f |∅).
π(f 1) = P (f1) = .03;
π(f 2) = P (f2) = .97;
// Compute F ’s π values.
send_π_msg(F, A);
The call
3.2. PEARL’S MESSAGE-PASSING ALGORITHM
149
send_π_msg(B, A);
results in the following steps:
π A (b1) = π(b1) = .005;
π A (b2) = π(b2) = .995;
// B sends A a π message.
π(a1) = P (a1|b1, f 1)π A (b1)π A (f 1) + P (a1|b1, f 2)πA (b1)πA (f 2)
+ P (a1|b2, f 1)πA (b2)πA (f 1) + P (a1|b2, f2)πA (b2)π A (f2)
= (.992)(.005)(1) + (.99)(.005)(1)
+ (.2)(.995)(1) + (.003)(.995)(1) = .212;
π(a2) = P (a2|b1, f 1)π A (b1)π A (f 1) + P (a2|b1, f 2)πA (b1)πA (f 2)
+ P (a2|b2, f 1)πA (b2)πA (f 1) + P (a2|b2, f2)πA (b2)π A (f2)
= (.008)(.005)(1) + (.01)(.005)(1)
+ (.8)(.995)(1) + (.997)(.995)(1) = 1.788;
P (a1|∅) = αλ(b1)π(b1) = α(1)(.202) = .212α;
P (a2|∅) = αλ(b2)π(b2) = α(1)(2.788) = 1.788α;
// Compute P (a|∅).
// This will not be
P (a1|∅) =
.212α
.212α+1.788α
= .106;
// P (a|∅) until A
P (a1|∅) =
1.788α
.212α+1.788α
= .894;
// gets F ’s π message.
The call
send_π_msg(F, A);
results in the following steps:
π A (f 1) = π(f 1) = .03;
π A (f 2) = π(f 2) = .97;
// F sends A a π
// message.
π(a1) = P (a1|b1, f 1)π A (b1)π A (f 1) + P (a1|b1, f 2)πA (b1)πA (f 2)
+ P (a1|b2, f 1)πA (b2)πA (f 1) + P (a1|b2, f2)πA (b2)π A (f2)
= (.992)(.005)(03) + (.99)(.005)(.97)
+ (.2)(.995)(03) + (.003)(.995)(.97) = .014;
π(a2) = P (a2|b1, f 1)π A (b1)π A (f 1) + P (a2|b1, f 2)πA (b1)πA (f 2)
+ P (a2|b2, f 1)πA (b2)πA (f 1) + P (a2|b2, f2)πA (b2)π A (f2)
= (.008)(.005)(.03) + (.01)(.005)(.97)
+ (.8)(.995)(.03) + (.997)(.995)(.97) = .986;
150
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
P (a1|∅) = αλ(b1)π(b1) = α(1)(.014) = .014α;
P (a2|∅) = αλ(b2)π(b2) = α(1)(.986) = .986α;
P (a1|∅) =
.014α
.014α+.986α
= .014;
P (a1|∅) =
.986α
.014α+.986α
= .986;
// Compute P (a|∅).
The initialized network is shown in Figure 3.10 (b).
Example 3.7 Consider again the Bayesian network in Figure 3.10 (a). Suppose A is instantiated for a1. That is, Antonio hears his burglar alarm sound.
Next we show the steps in the algorithm when the network’s values are updated
according to this instantiation.
The call
update_tree((G, P ), A, a, A, a1);
results in the following steps:
A = ∅ ∪ {A} = {A};
a = ∅ ∪ {a1} = {a1};
λ(a1) = 1; π(a1) = 1; P (a1|{a1}) = 1;
λ(a2) = 0; π(a2) = 0; P (a2|{a1}) = 0;
// Instantiate A for a1.
send_λ_msg(A, B);
send_λ_msg(A, F );
The call
send_λ_msg(A, B);
results in the following steps:
λA (b1) = [P (a1|b1, f 1)πA (f 1) + P (a1|b1, f2)π A (f2)] λ(a1)
= [P (a2|b1, f 1)πA (f 1) + P (a2|b1, f2)πA (f2)] λ(a2)
= [(.992)(.03) + (.99)(.97] 1 + [(.008)(.03) + (.01)(.97] 0
= .990;
// A sends B a λ message.
λA (b2) = [P (a1|b2, f 1)πA (f 1) + P (a1|b2, f2)π A (f2)] λ(a1)
= [P (a2|b2, f 1)πA (f 1) + P (a2|b2, f2)πA (f2)] λ(a2)
= [(.2)(.03) + (.003)(.97] 1 + [(.8)(.03) + (.997)(.97] 0
= .009;
3.2. PEARL’S MESSAGE-PASSING ALGORITHM
λ(b1) = λA (b1) = .990;
λ(b2) = λA (b2) = .009;
151
// Compute B’s λ values.
P (b1|{a1}) = αλ(b1)π(b1) = α(.990)(.005) = .005α;
P (b2|{a1}) = αλ(b2)π(b2) = α(.009)(.995) = .009α;
P (b1|{a1}) =
.005α
.005α+.0009α
= .357;
P (b2|{a1}) =
.009α
.005α+.0009α
= .643;
.
// Compute P (b|{a1}).
The call
send_λ_msg(A, F );
results in the following steps:
λA (f 1) = [P (a1|b1, f 1)πA (b1) + P (a1|b2, f1)πA (b2)] λ(a1)
= [P (a2|b1, f 1)πA (b1) + P (a2|b2, f1)πA (b2)] λ(a2)
= [(.992)(.005) + (.2)(.995)] 1 + [(.008)(.005) + (.8)(.995)] 0
= .204;
// A sends F a λ message.
λA (f 2) = [P (a1|b1, f 2)πA (b1) + P (a1|b2, f2)πA (b2)] λ(a1)
= [P (a2|b1, f 2)πA (b1) + P (a2|b2, f2)πA (b2)] λ(a2)
= [(.99)(.005) + (.003)(.995)] 1 + [(.01)(.005) + (.997)(.995] 0
= .008;
λ(f 1) = λA (f 1) = .204;
λ(f 2) = λA (f 2) = .008;
// Compute F ’s λ values.
P (f 1|{a1}) = αλ(f 1)π(f 1) = α(.204)(.03) = .006α;
P (f 2|{a1}) = αλ(f 2)π(f 2) = α(.008)(.97) = .008α;
P (f 1|{a1}) =
.006α
.008α+.006α
= .429;
P (f 2|{a1}) =
.008α
.008α+.006α
= .571;
.
// Compute P (f |{a1}).
The state of the network after this instantiation is shown in Figure 3.10 (c).
Notice the probability of a freight truck is greater than that of a burglar due to
the former’s higher prior probability.
Example 3.8 Consider again the Bayesian network in Figure 3.10 (a). Suppose after A is instantiated for a1, F is instantiated for f 1. That is, Antonio
sees a freight truck in back of his house. Next we show the steps in the algorithm
when the network’s values are updated according to this instantiation.
152
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
The call
update_tree((G, P ), A, a, F, f 1);
results in the following steps:
A = {A} ∪ {F } = {A, F };
a = {a1} ∪ {f 1} = {a1, f1};
λ(f 1) = 1; π(f 1) = 1; P (f 1|{f 1}) = 1;
λ(f 2) = 0; π(f 2) = 0; P (f 2|{f 1}) = 0;
// Instantiate F for f 1.
send_π_msg(F, A);
The call
send_π_msg(F, A);
results in the following steps:
πA (f 1) = π(f 1) = 1;
πA (f 2) = π(f 2) = 0;
// F sends A a π message.
send_λ_message(A, B);
The call
send_λ_msg(A, B);
results in the following steps:
λA (b1) = [P (a1|b1, f 1)πA (f 1) + P (a1|b1, f2)π A (f2)] λ(a1)
= [P (a2|b1, f 1)πA (f 1) + P (a2|b1, f2)πA (f2)] λ(a2)
= [(.992)(1) + (.99)(0)] 1 + [(.008)(1) + (.01)(0)] 0
= .992;
// A sends B a λ message.
λA (b2) = [P (a1|b2, f 1)πA (f 1) + P (a1|b2, f2)π A (f2)] λ(a1)
= [P (a2|b2, f 1)πA (f 1) + P (a2|b2, f2)πA (f2)] λ(a2)
= [(.2)(1) + (.003)(0)] 1 + [(.8)(.03) + (.997)(.97] 0
= .2;
λ(b1) = λA (b1) = .992;
λ(b2) = λA (b2) = .2;
// Compute B’s λ values.
P (b1|{a1, f 1}) = αλ(b1)π(b1) = α(.992)(.005) = .005α;
P (b2|{a1, f 1}) = αλ(b2)π(b2) = α(.2)(.995) = .199α;
3.2. PEARL’S MESSAGE-PASSING ALGORITHM
P (b1|{a1, f 1}) =
.005α
.005α+.199α
= .025;
P (b2|{a1, f 1}) =
.199α
.005α+.199α
= .975;
153
// Compute P (b|{a1, f 1}).
The state of the network after this instantiation is shown in Figure 3.10 (d).
Notice the discounting. The probability of a burglar drops from .357 to .025
when Antonio sees a freight truck in back of his house. However, since the two
causes are not mutually exclusive conditional on the alarm, it does not drop to
0. Indeed, it does not even drop to its prior probability .005.
3.2.3
Inference in Multiply-Connected Networks
So far we have considered only singly-connected networks. However, clearly
there are real applications in which this is not the case. For example, recall the
Bayesian network in Figure 3.1 is not singly-connected. Next we show how to
handle multiply-connected using the algorithm for singly-connected networks.
The method we discuss is called conditioning.
We illustrate the method with an example. Suppose we have a Bayesian
network containing a distribution P , whose DAG is the one in Figure 3.11
(a), and each random variable has two values. Algorithm 3.2 is not directly
applicable because the network is multiply-connected. However, if we remove
X from the network, the network becomes singly connected. So we construct
two Bayesian network, one of which contains the conditional distribution P 0 of P
given X = x1 and the other contains the conditional distribution P 00 of P given
X = x2. These networks are shown in Figures 3.11( b) and (c) respectively.
First we determine the conditional probability of every node given its parents
for each of these network. In this case, these conditional probabilities are the
same as the ones in our original network except for the roots Y and Z. For
those we have
P 0 (z1) = P (z1|x1)
P 0 (y1) = P (y1|x1)
P 00 (y1) = P (y1|x2)
P 0 (z1) = P (z1|x2).
We can then do inference in our original network by using Algorithm 3.2 to do
inference in each of these singly-connected networks. The following examples
illustrate the method.
Example 3.9 Suppose U is instantiated for u1 in the network in Figure 3.11
(a) . For the sake of illustration, consider the conditional probability of W given
this instantiation. We have
P (w1|u1) = P (w1|x1, u1)P (x1|u1) + P (w1|x2, u1)P (x2|u1).
The values of P (w1|x1, u1) and P (w1|x2, u1) can be obtained by applying Algorithm 3.2 to the networks in Figures 3.11( b) and (c) respectively. The value of
P (xi|u1) is given by
P (xi|u1) = αP (u1|xi)P (xi),
154
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
X = x1
X
P'(y1) = P(y1|x1)
X = x2
P'(z1) = P(z1|x1)
P''(y1) = P(y1|x2) P''(z1) = P(z1|x2)
Y
Z
Y
Z
Y
Z
W
T
W
T
W
T
U
U
U
(a)
(b)
(c)
Figure 3.11: A multiply-connected network is shown in (a). The singlyconnected networks obtained by instantiating X for x1 and for x2 are shown in
(b) and (c) respectively.
where is α a normalizing constant equal to 1/P (u1). The value of P (xi) is stored
in the network since X is a root, and the value of P (u1|xi) can be obtained by
applying Algorithm 3.2 to the networks in Figures 3.11( b) and (c). Thereby,
we can obtain the value of P (w1|u1). In the same way, we can obtain the
conditional probabilities of all non-conditioning variables in the network. Note
that along the way we have already computed the conditional probability (namely,
P (xi|u1)) of the conditioning variable.
Example 3.10 Suppose U is instantiated for u1 and Y is instantiated for y1
in the network in Figure 3.11 (a). We have
P (w1|u1, y1) = P (w1|x1, u1, y1)P (x1|u1, y1) + P (w1|x2, u1, y1)P (x2|u1, y1).
The values of P (w1|x1, u1, y1) and P (w1|x2, u1, y1) can be obtained by applying Algorithm 3.2 to the networks in Figures 3.11( b) and (c). The value of
P (xi|u1, y1) is given by
P (xi|u1, y1) = αP (u1, y1|xi)P (xi),
1
. The value of P (xi) is stored
where is α a normalizing constant equal to P (u1,y1)
in the network since X is a root. The value of P (u1, y1|xi) cannot be computed
3.2. PEARL’S MESSAGE-PASSING ALGORITHM
155
directly using Algorithm 3.2. But the chain rule enables us to obtain it with that
algorithm. That is,
P (u1, y1|xi) = P (u1|y1, xi)P (y1|xi).
The values on the right in this equality can both be obtained by applying Algorithm 3.2 to the networks in Figures 3.11( b) and (c).
The set of nodes, on which we condition, is called a loop-cutset. It is not
always possible to find a loop-cutset which contains only roots. Figure 3.16
in Section 3.6 shows a case in which we cannot. [Suermont and Cooper, 1990]
discuss criteria, which must be satisfied by the conditioning nodes, and they
present a heuristic algorithm for finding a set of nodes which satisfy these criteria. Furthermore, they prove the problem of finding a minimal loop-cutset is
N P -hard.
The general method is as follows. We first determine a loop-cutset C. Let E
be a set of instantiated nodes, and let e be their set of instantiations. Then for
each X ∈ V − {E ∪ C}, we have
X
P (xi|e, c)P (c|e),
P (xi) =
c
where the sum is over all possible values of the variables in C. The values of
P (xi|e, c) are computed using Algorithm 3.2. We determine P (c|e) using this
equality:
P (c|e) = αP (e|c)P (c).
To compute P (e|c) we first applying the chain as follows. If e = {e1 , ..., ek ),
P (e|c) = P (ek |ek−1 , ek−2 , ...e1 , c)P (ek−1 |ek−2 , ...e1 , c) · · · P (e1 |c).
Then Algorithm 3.2 is used repeatedly to compute the terms in this product.
The value of P (c) is readily available if all nodes in C are roots. As mentioned
above, in general, the loop-cutset does not contain only roots. A way to compute
P (c) in the general case is developed in [Suermondt and Cooper, 1991].
Pearl [1988] discusses another method for extending Algorithm 3.2 to handle
multiply-connected network called clustering.
3.2.4
Complexity of the Algorithm
Next we discuss the complexity of the algorithm. Suppose first the network is
a tree. Let
n =
k =
the number of nodes in the tree.
the maximum number of values for a node.
Then there are n−1 edges. We need to store at most k2 conditional probabilities
at each node, two k-dimensional vectors (the π and λ values) at each node, and
156
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
two k-dimensional vectors (the π and λ messages) at each edge. Therefore, an
upper bound on the number of values stored in the tree is
n(k2 + 2k) + 2(n − 1)k ∈ θ(nk2 ).
Let
c = maximum number of children over all nodes.
Then at most the number of multiplications needed to compute the conditional
probability of a variable is k to compute the π message, k2 to compute the
λ message, k2 to compute the π value, kc to compute the λ value, and k to
compute the conditional probability. Therefore, an upper bound on the number
of multiplications needed to compute all conditional probabilities is
¢
¡
n 2k2 + 2k + kc ∈ θ(nk2 + nkc).
It is not hard to see that, if a singly-connected network is sparse (i.e. each node
does not have many parents), the algorithm is still efficient in terms of space
and time. However, if a node has many parents, the space complexity alone
becomes intractable. In the next section, we discuss this problem and present a
model that solves it under certain assumptions. In Section 3.6, we discuss the
complexity in multiply-connected networks.
3.3
The Noisy OR-Gate Model
Recall that a Bayesian network requires the conditional probabilities of each
variable given all combinations of values of its parents. So, if each variable has
only two values, and a variable has p parents, we must specify 2p conditional
probabilities for that variable. If p is large, not only does our inference algorithm
become computationally unfeasible, but the storage requirements alone become
unfeasible. Furthermore, even if p is not large, the conditional probability of
a variable given a combination of values of its parents is ordinarily not very
accessible. For example, consider the Bayesian network in Figure 3.1 (shown at
the beginning of this chapter). The conditional probability of fatigue, given both
lung cancer and bronchitis are present, is not as accessible as the conditional
probabilities of fatigue given each is present by itself. Yet we need to specify this
former probability. Next we develop a model which requires that we need only
specify the latter probabilities. Not only are these probabilities more accessible,
but there are only a linear number of them. After developing the model, we
modify Algorithm 3.2 to execute efficiently using the model.
3.3.1
The Model
This model, called the noisy OR-gate model, concerns the case where the
relationships between variables ordinarily represent causal mechanism, and each
variable has only two values. The variable takes its first value if the condition
is present and its second value otherwise. Figure 3.1 illustrates such a case.
3.3. THE NOISY OR-GATE MODEL
157
For example, B (bronchitis) takes value b1 if bronchitis present and value b2
otherwise. For the sake of notational simplicity, in this section we show the
values only as 1 and 2. So B would take value 1 if bronchitis were present and
2 otherwise.
We make the following three assumptions in this model:
1. Causal inhibition: This assumption entails that there is some mechanism which inhibits a cause from bringing about its effect, and the presence of the cause results in the presence of the effect if and only if this
mechanism is disabled (turned off).
2. Exception independence: This assumption entails that the mechanism
that inhibits one cause is independent of the mechanism that inhibits
another causes.
3. Accountability: This assumption entails that an effect can happen only
if at least one of its causes is present and is not being inhibited. Therefore,
all causes which are not stated explicitly must be lumped into one unknown
cause.
Example 3.11 Consider again Figure 3.1. Bronchitis (B) and lung cancer
(C) both cause fatigue (F ). Causal inhibition implies that bronchitis will result
in fatigue if and only if the mechanism, that inhibits this from happening, is
not present. Exception independence implies that the mechanism that inhibits
bronchitis from resulting in fatigue behaves independently of the mechanism that
inhibits lung cancer form resulting in fatigue. Since we have listed no other
causes of fatigue in that figure, accountability implies fatigue cannot be present
unless at least one of bronchitis or lung cancer is present. Clearly, to use this
model in this example, we would have to add a third cause in which we lumped
all other causes of fatigue.
Given the assumptions in this model, the relationships among the variables
can be represented by the Bayesian network in Figure 3.12. That figure shows
the situation where there are n causes X1 , X2 , ... and Xn of Y . The variable
Ij is the mechanism that inhibits Xj . The Ij ’s are independent owing to our
assumption of exception independence. The variable Aj is on if and only if Xj
is present (equal to 1) and is not being inhibited. Owing to our assumption of
causal inhibition, this means Y should be present (equal to 1) if any one of the
Aj ’s is present. Therefore, we have
P (Y = 2|Aj = ON for some j) = 0.
This is why it called an ‘OR-gate’ model. That is, we can think of the Aj ’s
entering an OR-gate, whose exit feeds into Z (It is called ‘noisy’ because of the
Ij ’s). Finally, the assumption of accountability implies we have
P (Y = 2|A1 = OFF,A2 = OFF,...An = OFF) = 1.
We have the following theorem:
158
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
P(I1=ON) = q1
I1
P(In=ON) = qn
X1
In
Xn
P(A1=ON| I1=OFF,X1=1) = 1
P(A 1=ON| I1=OFF,X 1=2) = 0
P(An=ON| In=OFF,Xn=1) = 1
A1
An
P(An=ON| In=OFF,Xn=2) = 0
P(A 1=ON| I1=ON,X 1=1) = 0
P(An=ON| In=ON,Xn=1) = 0
P(A 1=ON| I1=ON,X 1=2) = 0
P(An=ON| In=ON,Xn=2) = 0
Y
P(Y=2|A1=OFF,A 2=OFF,...An=OFF) = 1
P(Y=2|A j=ON for some j) = 0
Figure 3.12: A Bayesian network representing the assumptions in the noisy
OR-gate model.
Theorem 3.3 Suppose we have a Bayesian network representing the Noisy Orgate model (i.e. Figure 3.12). Let
W = {X1 , X2 , ...Xn },
and let
w = {x1 , x2 , ...xn }
be a set of values of the variables in W. Furthermore, let S is a set of indices
such j ∈ S if and only if Xj = 1. That is,
S = {j such that Xj = 1}.
Then
P (Y = 2|W = w) =
Y
j∈S
Proof. We have
qj .
3.3. THE NOISY OR-GATE MODEL
159
P (Y = 2|W = w)
X
=
P (Y = 2|A1 = a1 , ...An = an )P (A1 = a1 , ...An = an |W = w)
a1 ,...an
=
X
P (Y = 2|A1 = a1 , ...An = an )
a1 ,...an
=
Y
Y
j
P (Aj = aj |Xj = xj )
P (Aj = OFF|Xj = xj )
j
=
Y
[P (Aj = OFF|Xj = xj , Ij = ON)P (Ij = ON) +
j
P (Aj = OFF|Xj = xj , Ij = OFF)P (Ij = OFF)]



Y
Y
=  1(qj ) + 1(1 − qj )  1(qj ) + 0(1 − qj )

= 
j ∈S
/
Y
j ∈S
/

1 
Y
j∈S

qj  =
j∈S
Y
qj .
j∈S
Our actual Bayesian network contains Y and the Xj ’s but it does not contain
the Ij ’s or Aj ’s. In that network, we need to specify the conditional probability
of Y given each combination of values of the Xj ’s. Owing to the preceding
theorem, we need only specify the values of qj for all j. All necessary conditional
probabilities can then be computed using Theorem 3.3. Instead, we often specify
pj = 1 − qj ,
which is called the causal strength of X for Y . Theorem 3.3 implies
pj = P (Y = 1|Xj = 1, Xi = 2 for i 6= j).
This value is relatively accessible. For example, we may have a reasonably large
database of patients, whose only disease is lung cancer. To estimate the causal
strength of lung cancer for fatigue, we need only determine how many of these
patients are fatigued. On the other hand to directly estimate the conditional
probability of fatigue given lung cancer, bronchitis, and other causes, we would
need databases containing patients with all combinations of these diseases.
Example 3.12 Suppose we have the Bayesian network in Figure 3.13, where
the causal strengths are shown on the edges. Owing to Theorem 3.3,
P (Y = 2|X1 = 1, X2 = 2, X3 = 1, X4 = 1) = (1 − p1 )(1 − p3 )(1 − p4 )
= (1 − .7)(1 − .6)(1 − .9)
= .012.
So
P (Y = 1|X1 = 1, X2 = 2, X3 = 1, X4 = 1) = 1 − .012 = .988.
160
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
X1
X2
p1 = .7
X3
p2 = .8
p3 = .6
X4
p4 = .9
Z
Figure 3.13: A Bayesian network using the Noisy OR-gate model.
3.3.2
Doing Inference With the Model
Even though Theorem 3.3 solves our specification problem, we still need to
compute possibly an exponential number of values to do inference using Algorithm 3.2. Next we modify that algorithm to do inference more efficiently with
probabilities specified using the noisy OR-gate model.
Assume the variables satisfy the noisy OR-gate model, and Y has n parents
X1 , X2 , ... and Xn . Let pj be the causal strength of Xj for Y , and qj = 1 − pj .
The situation with n = 4 is shown in Figure 3.13. Before proceeding, we alter
our notation a little. That is, to denote that Xj is present, we use x+
j instead
−
of 1; to denote that Xj is absent, we use xj instead of 2.
Consider first the λ messages. Using our present notation, we must do the
following computation in Algorithm 3.2 to calculate the λ message Y sends to
Xj :



X
X
Y

P (y|x1 , x2 , . . . xn )
π Y (xi ) λ(y).
λY (xj ) =
y
x1 ,...xj−1 ,xj+1 ,...xn
i6=j
We must determine an exponential number of conditional probabilities to do
this computation. It is left as an exercise to show that, in the case of the Noisy
OR-gate model, this formula reduces to the following formulas:
−
+
λY (x+
j ) = λ(y )qj Pj + λ(y )(1 − qj Pj )
(3.4)
−
+
λY (x−
j ) = λ(y )Pj + λ(y )(1 − Pj )
(3.5)
where
Pj =
Y
[1 − pi π Y (x+
i )].
i6=j
Clearly, this latter computation only requires that we do a linear number of
operations.
3.4. OTHER ALGORITHMS THAT EMPLOY THE DAG
161
Next consider the π values. Using our present notation, we must do the
following computation in Algorithm 3.2 to compute the π value of Y :


n
Y
X
P (y|x1 , x2 , . . . xn )
πY (xj )
π(y) =
x1 ,x2 ,...xn
j=1
We must determine an exponential number of conditional probabilities to do
this computation. It is also left as an exercise to show that, in the case of the
Noisy OR-gate model, this formula reduces to the following formulas:
π(y + ) = 1 −
π(y− ) =
n
Y
j=1
n
Y
[1 − pj πY (x+
j )]
[1 − pj πY (x+
j )].
(3.6)
(3.7)
j=1
Again, this latter computation only requires that we do a linear number of
operations.
3.3.3
Further Models
A generalization of the Noisy OR-gate model to the case of more than two
values appears in [Srinivas, 1993]. Other models for succinctly representing the
conditional distributions include the sigmoid function (See [Neal, 1992].) and
the logit function (See [McLachlan and Krishnan, 1997].) Another approach to
reducing the number of parameter estimates is the use of embedded Bayesian
networks, which is discussed in [Heckerman and Meek, 1997]. Note that their
use of the term ‘embedded Bayesian network’ is different than our use in Chapter
6.
3.4
Other Algorithms that Employ the DAG
Shachter [1988] created an algorithm which does inference by performing arc
reversal/node reduction operations in the DAG. The algorithm is discussed
briefly in Section 5.2.2; however, you are referred to the original source for a
detailed discussion.
Based on a method originated in [Lauritzen and Spiegelhalter, 1988], Jensen
et al [1990] developed an inference algorithm that involves the extraction of an
undirected triangulated graph from the DAG in a Bayesian network, and the
creation of a tree whose vertices are the cliques of this triangulated graph. Such
a tree is called a junction tree.. Conditional probabilities are then computed
by passing messages in the junction tree. You are referred to the original source
and to [Jensen, 1996] for a detailed discussion of this algorithm, which we call
the Junction tree Algorithm.
162
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
X
Y
Z
W
T
Figure 3.14: A DAG.
3.5
The SPI Algorithm
The algorithms discussed so far all do inference by exploiting the conditional
independencies entailed by the DAG. Pearl’s method does this by passing messages in the original DAG, while Jensen’s method does it by passing messages in
the junction tree obtained from the DAG. D’Ambrosio and Li [1994] took a different approach. They developed an algorithm which approximates finding the
optimal way to compute marginal distributions of interest from the joint probability distribution. They call this symbolic probabilistic inference (SPI).
First we illustrate the method with an example.
Suppose we have a joint probability distribution determined by conditional
distributions specified for the DAG in Figure 3.14 and all variables are binary.
Then
P (x, y, z, w, t) = P (t|z)P (w|y, z)P (y|x)P (z|x)P (x).
Suppose further we wish to compute P (t|w) for all values of T and W . We have
P
P (t, w)
x,y,z P (x, y, z, w, t)
= P
P (t|w) =
P (w)
x,y,z,t P (x, y, z, w, t)
P
x,y,z P (t|z)P (w|y, z)P (y|x)P (z|x)P (x)
.
= P
x,y,z,t P (t|z)P (w|y, z)P (y|x)P (z|x)P (x)
To compute the sums in the numerator and denominator of the last expression
by the brute force method of individually computing all terms and adding them
is computationally
very costly. For specific values of T and W we would have
¡ ¢
to do 23 4 = 32 multiplications to compute the sum in the numerator. Since
there are four combinations of values of T and W , this means we would have
have to do 128 multiplications to compute all numerators. We can save time by
not re-computing a product each time it is needed. For example, suppose we do
3.5. THE SPI ALGORITHM
163
the multiplications in the order determined by the factorization that follows:
X
P (t, w) =
[[[[P (t|z)P (w|y, z)] P (y|x)] P (z|x)] P (x)]
(3.8)
x,y,z
The first product involves 4 variables, which means 24 multiplications are required to compute its value for all combinations of the variables; the second,
third and fourth products each involve 5 variables, which means 25 multiplications are required for each. So the total number of multiplications required is
112, which means we saved 16 multiplications by not recomputing products.
We can save more multiplications by summing over a variable once it no longer
appears in remaining terms. Equality 3.8 then becomes
##
"
"
X
X
X
[[P (t|z)P (w|y, z)] P (y|x)] .
(3.9)
P (x)
P (z|x)
P (t, w) =
x
z
y
The first product again involves 4 variables and requires 24 multiplications, and
the second again involves 5 variables and requires 25 multiplications. However,
we sum y out before taking the third product. So it involves only 4 variables
and requires 24 multiplications. Similarly, we sum z out before taking the fourth
product, which means it only involves 3 variables and requires 23 multiplications.
Therefore, the total number of multiplications required is only 72.
Different factorizations can require different numbers of multiplications. For
example, consider the factorization that follows:
##
"
"
X
X
X
[P (y|x) [P (z|x)P (x)]] .
(3.10)
P (t|z)
P (w|y, z)
P (t, w) =
z
y
x
It is not hard to see that this factorization requires only 28 multiplications.
To minimize the computational effort involved in computing a given marginal
distribution, we want to find the factorization that requires the minimal number
of multiplications. D’Ambrosio and Li [1994] called this the Optimal factoring
Problem. They formulated the problem for the case of binary variables (There
is a straightforward generalization to multinomial variables.). After developing
the formalization, we apply it to probabilistic inference.
3.5.1
The Optimal Factoring Problem
We start with a definition.
Definition 3.1 A factoring instance F = (V, S, Q) consists of
1. a set V of size n;
ª
©
2. A set S of m subsets S{1} , . . . S{m} of V;
3. A subset Q ⊆ V called the target set.
164
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
Example 3.13 The following is a factoring instance:
1. n = 5 and V = {x, y, z, w, t}.
2. m = 5 and
S{1}
S{2}
S{3}
S{4}
S{5}
=
=
=
=
=
{x}
{x, z}
{x, y}
{y, z, w}
{z, t}.
3. Q = {w, t}.
ª
©
Definition 3.2 Let S = S{1} , . . . S{m} . A factoring α of S is a binary tree
with the following properties:
1. All and only the members of S are leaves in the tree.
2. The parent of nodes SI and SJ is denoted SI∪J .
3. The root of the tree is S{1,...m} .
We will apply factorings to factoring instances. However, note that a factoring is independent of the actual values of the S{i} in a factoring instance.
ª
©
Example 3.14 Suppose S = S{1} , . . . S{5} . Then three factorings of S appear
in Figure 3.15.
Given a factoring instance F = (V, S, Q) and a factoring α of S, we compute
the cost µα (F) as follows. Starting at the leaves of α, we compute the values
of all nodes according to this formula:
SI∪J = SI ∪ SJ − WI∪J
where
¢
ª
©
¡
/ Q) .
/ I ∪ J, v ∈
/ S{k} and (v ∈
WI∪J = v : for all k ∈
As the nodes’ values are determined, we compute the cost of the nodes according
to this formula:
and
¡
¢
µα S{j} = 0
for
1≤j≤m
µα (SI∪J ) = µα (SI ) + µα (SJ ) + 2|SI ∪SJ | ,
where || is the number of elements in the set. Finally, we set
¡
¢
µα (F) = µα S{1,...m} .
3.5. THE SPI ALGORITHM
165
S{1,2,3,4,5}
S{1,2,3,4,5}
S{1,2,3,4}
S{1,2,3}
S{1,2}
S{1}
S{2,3,4,5}
S{5}
S{3,4,5}
S{4}
S{4,5}
S{3}
S{4}
S{2}
S{2}
S{3}
S{5}
(b)
(a)
S{1,2,3,4,5}
S{1,2}
S{1}
S{3,4,5}
S{2}
S{3,4}
S{3}
S{1}
S{5}
S{4}
(c)
ª
©
Figure 3.15: Three factorings of S = S{1} , . . . S{5} .
166
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
Example 3.15 Suppose we have the factoring instance F in Example 3.13.
Given the factoring α in Figure 3.15 (a), we have
S{1,2}
S{1,2,3}
S{1,2,3,4}
S{1,2,3,4,5}
= S{1} ∪ S{2} − W{1,2}
= {x} ∪ {x, z} − ∅ = {x, z}
= S{1,2} ∪ S{3} − W{1,2,3}
= {x, z} ∪ {x, y} − {x} = {y, z}
= S{1,2,3} ∪ S{4} − W{1,2,3,4}
= {y, z} ∪ {y, z, w} − {x, y} = {z, w}
= S{1,2,3,4} ∪ S{5} − W{1,2,3,4,5}
= {z, w} ∪ {z, t} − {x, y, z} = {w, t}.
Next we compute the cost:
¡
¢
¡
¢
¡
¢
µα S{1,2} = µα S{1} + µα S{2} + 22
= 0+0+4 = 4
¡
¢
¡
¢
¡
¢
µα S{1,2,3} = µα S{1,2} + µα S{3} + 23
= 4 + 0 + 8 = 12
¡
¢
¡
¢
¡
¢
µα S{1,2,3,4} = µα S{1,2,3} + µα S{4} + 23
= 12 + 0 + 8 = 20
¡
¢
¡
¢
¡
¢
µα S{1,2,3,4,5} = µα S{1,2,3,4} + µα S{5} + 23
= 20 + 0 + 8 = 28.
So
¡
¢
µα (F) = µα S{1,2,3,4,5} = 28.
Example 3.16 Suppose again we have the factoring instance F in Example
3.13. Given the factoring β in Figure 3.15 (b), we have
S{4,5}
S{3,4,5}
= S{4} ∪ S{5} − W{4,5}
= {y, z, w} ∪ {z, t} − ∅ = {y, z, w, t}
= S{4,5} ∪ S{3} − W{3,4,5}
= {y, z, w, t} ∪ {x, y} − {y} = {x, z, w, t}
3.5. THE SPI ALGORITHM
S{2,3,4,5}
167
= S{3,4,5} ∪ S{2} − W{2,3,4,5}
= {x, z, w, t} ∪ {x, z} − {y, z} = {x, w, t}
S{1,2,3,4.5}
= S{2,3,4,5} ∪ S{1} − W{1,2,3,4,5}
= {x, w, t} ∪ {x} − {x, y, z} = {w, t}.
It is left as an exercise to show
µβ (F) = 72.
Example 3.17 Suppose we have the following factoring instance F0 :
1. n = 5 and V = {x, y, z, w, t}.
2. m = 5 and
S{1}
S{2}
S{3}
S{4}
S{5}
=
=
=
=
=
{x}
{y}
{z}
{w}
{x, y, z, w, t}.
3. Q = {t}.
Given the factoring γ in Figure 3.15 (c), we have
S{3,4,5}
S{1,2}
= S{1} ∪ S{2} − W{1,2}
= {x} ∪ {y} − ∅ = {x, y}
S{3,4}
= S{3} ∪ S{4} − W{3,4}
= {z} ∪ {w} − ∅ = {z, w}
= S{3,4} ∪ S{5} − W{3,4,5}
= {z, w} ∪ {x, y, z, w, t} − {z, w} = {x, y, t}
S{1,2,3,4,5}
= S{1,2} ∪ S{3,4,5} − W{1,2,3,4,5}
= {x, y} ∪ {x, y, t} − {x, y, z, w} = {t}.
Next we compute the cost:
¡
¢
¡
¢
¡
¢
µγ S{1,2} = µγ S{1} + µγ S{2} + 22
= 0+0+4 = 4
168
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
¡
¢
¡
¢
¡
¢
µγ S{3,4} = µγ S{3} + µγ S{4} + 22
= 0+0+4 =4
¡
¢
¡
¢
¡
¢
µγ S{3,4,5} = µγ S{3,4} + µγ S{5} + 25
= 4 + 0 + 32 = 36
¡
¢
¡
¢
¡
¢
µγ S{1,2,3,4,5} = µγ S{1,2} + µγ S{3,4,5} + 23
= 4 + 36 + 8 = 48.
So
¡
¢
µγ (F0 ) = µγ S{1,2,3,4,5} = 48.
Example 3.18 Suppose we have the factoring instance F0 in Example 3.17. It
is left as an exercise to show for the factoring β in Figure 3.15 (b) that
µβ (F0 ) = 60.
We now state the Optimal factoring Problem. Namely, the Optimal factoring Problem is to find a factoring α for a factoring instance F such that
µα (F) is minimal.
3.5.2
Application to Probabilistic Inference
Notice that the cost µα (F), computed in Example 3.15, is equal to the number
of multiplications required by the factorization in Equality 3.10; and the cost
µβ (F), computed in Example 3.16, is equal to the number of multiplications
required by the factorization in Equality 3.9. This is no coincidence. We can
associate a factoring instance with every marginal probability computation in a
Bayesian network, and any factoring of the set S in the instance corresponds to
a factorization for the computation of that marginal probability. We illustrate
the association next. Suppose we have the Bayesian network in Figure 3.14.
Then
P (x, y, z, w, t) = P (t|z)P (w|y, z)P (y|x)P (z|x)P (x).
Suppose further that as before we want to compute P (w, t) for all values of W
and T . The factoring instance corresponding to this computation is the one
shown in Example 3.13. Note that there is an element in S for each conditional
probability expression in the product, and the members of an element are the
variables in the conditional probability expression. Suppose we compute P (w, t)
using the factorization in Equality 3.10, which we now show again:
##
"
"
X
X
X
[P (y|x) [P (z|x)P (x)]] .
P (t|z)
P (w|y, z)
P (t, w) =
z
y
x
3.5. THE SPI ALGORITHM
169
The factoring α in Figure 3.15 (a) corresponds to this factorization. Note that
the partial order in α of the subsets is the partial order in which the corresponding conditional probabilities are multiplied. Similarly, the factoring β in Figure
3.15 (b) corresponds to the factorization in Equality 3.9.
D’Ambrosio and Li [1994] show that, in general, if F is the factoring instance
corresponding to a given marginal probability computation in a Bayesian network, then the cost µα (F) is equal to the number of multiplications required by
the factorization to which α corresponds. So if we solve the Optimal factoring
Problem for a given factoring instance, we have found a factorization which
requires a minimal number of multiplications for the marginal probability computation to which the factoring instance corresponds. They note that each
graph-based inference algorithms corresponds to a particular factoring strategy.
However, since a given strategy is constrained by the structure of the original
DAG (or of a derived junction tree), it may be hard for the strategy to find an
optimal factoring.
D’Ambrosio and Li [1994] developed a linear time algorithm which solves the
Optimal factoring Problem when the DAG in the corresponding Bayesian network is singly-connected. Furthermore, they developed a θ(n3 ) approximation
algorithm for the general case.
The total computational cost when doing probabilistic inference using this
technique includes the time to find the factoring (called symbolic reasoning)
and the time to compute the probability (called numeric computation). The
algorithm for doing probabilistic inference, which consists of both the symbolic
reasoning and the numeric computation, is called the Symbolic probabilistic
inference (SPI) Algorithm.
The Junction tree Algorithm is considered overall to be the best graph-based
algorithm (There are, however, specific instances in which Pearl’s Algorithm is
more efficient. See [Neapolitan, 1990] for examples.). If the task is to compute all marginals given all possible sets of evidence, it is believed one cannot
improve on the Junction tree Algorithm (ignoring factorable local dependency
models such as the noisy OR-gate model). However, even that has never been
proven. Furthermore, it seems to be a somewhat odd problem definition. For
any specific pattern of evidence, one can often do much better than the generic
evidence-independent junction tree. D’Ambrosio and Li [1994] compared the
performance of the SPI Algorithm to the Junction tree Algorithm using a number of different Bayesian networks and probability computations, and they found
that the SPI Algorithm performed dramatically fewer multiplications. Furthermore, they found the time spent doing symbolic reasoning by the SPI Algorithm
was insignificant compared to the time spent doing numeric computation.
Before closing, we note that SPI is not the same as simply eliminating variables as early as possible. The following example illustrates this:
Example 3.19 Suppose our joint probability distribution is
P (t|x, y, z, w)P (w)P (z)P (y)P (x),
and we want to compute P (t) for all values of T . The factoring instance F0 in
170
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
Example 3.17 corresponds to this marginal probability computation. The following factorization eliminates variables as early as possible:
###
"
"
"
X
X
X
X
[P (t|x, y, z, w)P (w)]
.
P (x)
P (y)
P (z)
x
y
z
w
The factoring β in Figure 3.15 (b) corresponds to this factorization. As shown
in Example 3.18 µβ (F0 ) = 60, which means this factorization requires 60 multiplications.
On the other hand, consider this factorization:
#
"
XX
XX
[P (t|x, y, z, w) [P (w)P (z)]] .
[P (x)P (y)]
y
x
z
w
The factoring γ in Figure 3.15 (c) corresponds to this factorization. As shown
in Example 3.17 µγ (F0 ) = 48, which means this factorization requires only 48
multiplications.
Bloemeke and Valtora [1998] developed a hybrid algorithm based on the
junction tree and symbolic probabilistic methods.
3.6
Complexity of Inference
First we show that using conditioning and Algorithm 3.2 to handle inference
in a multiply-connected network can sometimes be computationally unfeasible.
Suppose we have a Bayesian network, whose DAG is the one in Figure 3.16.
Suppose further each variable has two values. Let k be the depth of the DAG.
In the figure, k = 6. Using the method of conditioning presented in Section
3.2.3, we must condition on k/2 nodes to render the DAG singly connected.
That is, we must condition on all the nodes on the far left side or the far right
side of the DAG. Since each variable has two values, we must therefore perform
inference in θ(2k/2 ) singly-connected networks in order to compute P (y1|x1).
Although the Junction tree and SPI Algorithms are more efficient than
Pearl’s algorithm for certain DAGs, they too are worst-case non-polynomial
time. This is not surprising since the problem of inference in Bayesian networks
has been shown to be N P -hard. Specifically, [Cooper, 1990] has obtained the
result that, for the set of Bayesian networks that are restricted to having no
more than two values per node and no more than two parents per node, with no
restriction on the number of children per node, the problem of determining the
conditional probabilities of remaining variables given certain variables have been
instantiated, in multiply-connected networks, is #P -complete. #P -complete
problems are a special class of N P -hard algorithms. Namely, the answer to a
#P -complete problem is the number of solutions to some N P -complete problem.
In light of this result, researchers have worked on approximation algorithms for
inference Bayesian networks. We show one such algorithm in the next chapter.
3.7. RELATIONSHIP TO HUMAN REASONING
171
X
depth = 6
Y
Figure 3.16: Our method of conditioning will require exponential time to compute P (y1|x1).
3.7
Relationship to Human Reasoning
First we present the causal network model, which is a model of how humans
reason with causes. Then we show results of studies testing this model.
3.7.1
The Causal Network Model
Recall from Section 1.4 that if we identify direct causes-effect relationships
(edges) by any means whatsoever, draw a causal DAG using the edges identified, and assume the probability distribution of the variables satisfies the Markov
condition with this DAG, we are making the causal Markov assumption. We
argued that, when causes are identified using manipulation, we can often make
the causal Markov assumption, and hence the casual DAG, along with its conditional distributions, constitute a Bayesian network that pretty well models
reality. That is, we argued that relationships, which we objectively define as
causal, constitute a Bayesian network in external reality. Pearl [1986, 1995]
172
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
burglar
foorprints
earthquake
alarm
Figure 3.17: A causal network.
takes this argument a step further. Namely, he argues that a human internally
structures his or her causal knowledge in his or her personal Bayesian network,
and that he or she performs inference using that knowledge in the same way
as Algorithm 3.2. When the DAG in a Bayesian network is a causal DAG, the
network is called a causal network. Henceforth, we will use this term, and we
will call this model of human reasoning the causal network model. Pearl’s
argument is not that a globally consistent causal network exists at a cognitive
level in the brain. ‘Instead, fragmented structures of causal organizations are
constantly being assembled on the fly, as needed, from a stock of functional
building blocks’ - [Pearl, 1995].
Figure 3.17 shows a causal network representing the reasoning involved when
a Mr. Holmes learns that his burglar alarm has sounded. He knows that earthquakes and burglars can both cause his alarm to sound. So there are arcs from
both earthquake and burglar to alarm. Only a burglar could cause footprints to
be seen. So there is an arc only from burglar to footprints. The causal network
model maintains that Mr. Holmes reasons as follows. If he were in his office at
work and learned that his alarm had sounded at home, he would assemble the
cause-effect relationship between burglar and alarm. He would reason along the
arc from alarm to burglar to conclude that he had probably been burglarized. If
he later learned of an earthquake, he would assemble the earthquake-alarm relationship. He would then reason that the earthquake explains away the alarm,
and therefore he had probably not been burglarized. Notice that according to
this model, he mentally traces the arc from earthquake to alarm, followed by
the one from alarm to burglar. If, when Mr. Holmes got home, he saw strange
footprints in the yard, he would assemble the burglar-footprints relationship
and reason along the arc between them. Notice that this tracing of arcs in the
causal network is how Algorithm 3.2 does inference in Bayesian networks. The
causal network model maintains that a human reasons with a large number of
nodes by mentally assembling small fragments of causal knowledge in sequence.
The result of reasoning with the link assembled in one time frame is used when
reasoning in a future time frame. For example, the determination that he has
3.7. RELATIONSHIP TO HUMAN REASONING
173
probably been burglarized (when he learns of the alarm) is later used by Mr.
Holmes when he sees and reasons with the footprints.
Tests on human subjects have been performed testing the accuracy of the
causal network model. We discuss that research next.
3.7.2
Studies Testing the Causal Network Model
First we discuss ‘discounting’ studies, which did not explicitly state they were
testing the causal network model, but were doing so implicitly. Then we discuss
tests which explicitly tested it.
Discounting Studies
Psychologists have long been interested in how an individual judges the presence of a cause when informed of the presence of one of its effect, and whether
and to what degree the individual becomes less confident in the cause when
informed that another cause of the effect was present. Kelly [1972] called this
inference discounting. Several researchers ([Jones, 1979], [Quattrone, 1982],
[Einhorn and Hogarth, 1983], [McClure, 1989]) have argued that studies indicate that in certain situations people discount less than is warranted. On the
other hand, arguments that people discount more than is warranted also have a
long history (See [Mills, 1843], [Kanouse, 1972], and [Nisbett and Ross, 1980].).
In many of the discounting studies, individuals were asked to state their feelings about the presence of a particular cause when informed another cause was
present. For example, a classic finding is that subjects who read an essay defending Fidel Castro’s regime in Cuba ascribe a pro-Castro attitude to the essay
writer even when informed that the writer was instructed to take a pro-Castro
stance. Researchers interpreted these results as indicative of underdiscounting.
Morris and Larrick [1995] argue that the problem in these studies is that the
researchers assume that subjects believe a cause is sufficient for an effect when
actually the subjects do not believe this. That is, the researchers assumed the
subjects believed the probability is 1 that an effect is present conditional on
one of its causes being present. Morris and Larrick [1995] repeated the Castro
studies, but used subjective probability testing instead of assuming, for example, that the subject believes an individual will always write a pro-Castro essay
whenever told to do so (They found that subjects only felt it was highly probable
this would happen.). When they replaced deterministic relationships by probabilistic ones, they found that subjects discounted normatively. That is, using
as a benchmark the amount of discounting implied by applying Bayes’ rule, they
found that subjects discounted about correctly. Since the causal network model
implies subjects would reason normatively, their results support that model.
Plach’s Study
While research on discounting is consistent with the causal network model, the
inference problems considered in this research involved very simple networks
174
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
(e.g., one effect and two causes). One of the strengths of causal networks is
the ability to model complex relationships among a large number of variables.
Therefore, research was needed to examine whether human causal reasoning
involving more complex problems can be effectively modeled using a causal network. To this end, Plach [1997] examined human reasoning in larger networks
modeling traffic congestion. Participants were asked to judge the probability
of various traffic-related events (weather, accidents, etc.), and then asked to
update their estimate of the probability of traffic congestion as additional evidence was made available. The results revealed a high correspondence between
subjective updating and normative values implied by the network. However,
there were several limitations to this study. All analyses were performed on
probability estimates, which had been averaged across subjects. To the extent
that individuals differ in their subjective beliefs, these averages may obscure
important individual differences. Second, participants were only asked to consider two pieces of evidence at a time. Thus, it is unclear whether the result
would generalize to more complex problems with larger amounts of evidence.
Finally, participants were asked to make inferences from cause to effect, which
is distinct from the diagnostic task where inferences must be made from effects
to causes.
Morris and Neapolitan’s Study
Morris and Neapolitan [2000] utilized an approach similar to Plach’s to explore
causal reasoning in computer debugging. However, they examined individuals’
reasoning with more complex causal relationships and with more evidence. We
discuss their study in more detail.
Methodology First we give the methodology.
Participants The participants were 19 students in a graduate-level computer science course. All participants had some experience with the type of
program used in the study. Most participants (88%) rated their programming
skill as either okay or good, while the remainder rated their skill level as expert.
Procedure The study was conducted in three phases. In the first phase,
two causal networks were presented to the participants and discussed at length
to familiarize participants with the content of the problem. The causal networks
had been developed based on interviewing an experienced computer programmer
and observing him while he was debugging code. Both networks described
potential causes of an error in a computer program, which was described as
follows:
One year ago, your employer asked you to create a program to
verify and insert new records into a database. You finished the program and it compiled without errors. However, the project was put
on hold before you had a chance to fully test the program. Now, one
3.7. RELATIONSHIP TO HUMAN REASONING
Inappropriate
PID in data file
Error in Error Log
print statement
175
Program
alters PID
Inappropriate value
assigned to PID
Inappropriate PID
in error log
Figure 3.18: Causal network for a simple debugging problem.
year later, your boss wants you to implement the program. While
you remember the basic function of the program (described below),
you can’t recall much of the detail of your program. You need to
make sure the program works as intended before the company puts
it into operation.
The program is designed to take information from a data file (the
Input File) and add it to a database. The database is used to track
shipments received from vendors, and contains information relating
to each shipment (e.g., date of arrival, mode of transportation, etc.),
as well as a description of one or more packages within each shipment
(e.g., product type, count, invoice number, etc.). Each shipment is
given a unique Shipment Identification code (SID), and each package
is given a unique Package Identification code (PID).
The database has two relations (tables). The Shipment Table
contains information about the entire shipment, and the Package
Table contains information about individual packages. SID is the
primary key for the Shipment Table and a foreign key in the Package
Table. PID is the primary key for the Package Table.
If anything goes wrong with the insertion of new records (e.g.,
there are missing or invalid data), the program writes the key information to a file called the Error Log. This is not a problem as long
as records are being rejected because they are invalid. However, you
need to verify that errors are written correctly to the Error Log.
Two debugging tasks were described. The first problem was to determine
why inappropriate PID values were found in the Error Log. The causal network
for this problem was fairly simple, containing only five nodes (See Figure 3.18.).
The second problem was to determine why certain records were not added to the
database. The causal network for this problem was considerably more complex,
containing 14 variables (See Figure 3.19.).
In the second phase, participants’ prior beliefs about the events in each
network were measured. For events with no causes, participants were asked
to indicate the prior probability, which was defined as the probability of the
event occurring when no other information is known. For events that were
176
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
Program alters shipment
record SID (e.g., truncation)
Ship ment record
repeated in Input File.
SID for shipment
record in Input File
has invalid format
Program tried to insert two
records with the same SID
into the Ship ment Table
Error Message:
Primary key has field
with null key value.
Failed to add
shipment record to
Shipment Table
Wrong package
record SID value
in Input File
SID fro m package
record could not be
matched to a value in
the Shipment Tab le
Failed to add
package record to
Package Table
Wrong shipment
record SID value in
Input File
Error message:
Duplicate value
in unique key.
Wrong SID in
Shipment Table
Several package records
in the Error Log have
the same SID
Error message:
Vio lation of
Integrity Ru le 2
Figure 3.19: Causal network for a complex debugging problem.
caused by other events in the network, participants were asked to indicate the
conditional probabilities. Participants indicated the probability of the effect,
given that each cause was known to have occurred in isolation, assuming that
no other causes had occurred. In addition, participants rated the probability of
the effect occurring when none of the causes were present. From this data, all
conditional probabilities were computed using the noisy OR-gate model.
All probabilities were obtained using the method described in [Plach, 1997].
Participants were asked to indicate the number of times, out of 100, that an event
would occur. So probabilities were measured on a scale from 0 to 100. Examples
of both prior and conditional probabilities were presented to participants and
discussed to ensure that everyone understood the rating task.
In the third phase of the study, participants were asked to update the probabilities of events as they received evidence about the values of particular nodes.
Participants were first asked to ascertain the prior probabilities of the values of
every node in the network. They were then informed of the value of a particular
node, and they were asked to determine the conditional probabilities of the values of all other nodes given this evidence. Several pieces of additional evidence
were given in each block of trails. Four blocks of trials were conducted, two
involving the first network, and two involving the second network.
The following evidence was provided in each block:
1. Block 1 (refers to the network in Figure 3.18)
Evidence 1. You find an inappropriate PID in the error log.
Evidence 2. You find an error in the Error Log print statement.
3.7. RELATIONSHIP TO HUMAN REASONING
177
2. Block 2 (refers to the network in Figure 3.18)
Evidence 1. You find an inappropriate PID in the error log.
Evidence 2. You find that there are no inappropriate PIDs in the data
file.
3. Block 3 (refers to the network in Figure 3.19)
Evidence 1. You find there is a failure to add several package records to
the Package Table.
Evidence 2. You get the message ‘Error Message: Violation of integrity
rule 2.’
Evidence 3. You find that several package records in the Error Log have
the same SID.
Evidence 4. You get the message ‘Error Message: Duplicate value in
unique key.’
4. Block 4 (refers to the network in Figure 3.19)
Evidence 1. You find there is a failure to add a shipment record to the
Shipment Table.
Evidence 2. You get the message ‘Error Message: Primary key has field
with null key value.’
Statistical Analysis The first step in the analysis was to model participants’ subjective causal networks. A separate Bayesian network was developed
for each participant based on the subjective probabilities gathered in Phase 2.
Each of these networks was constructed using the Bayesian network inference
program, Hugin (See [Olesen et al, 1992].). Then nodes in the network were
instantiated using the same evidence as was provided to participants in Phase
3 of the study. The updated probabilities produced by Hugin were used as
normative values for the conditional probabilities.
The second step in the analysis was to examine the correspondence between
participants and the Bayesian networks, which was defined as the correlation between subjective and normative probabilities. In addition, the analysis included
an examination of the extent to which correspondence changed as a function of
1) the complexity of the network, 2) the amount of evidence provided, and 3)
the participant providing the judgements.
The correspondence between subjective and normative ratings was examined
using hierarchical linear model (HLM) analysis [Bryk, 1992]. The primary result
of interest was the determination of the correlation between normative and
subjective probabilities. These results are shown in Figure 3.20.
178
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
Correlation between normative and
subjective probability
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
1
2
3
4
Number of Pieces of Evidence Presented
Simple Network
Complex Network
Figure 3.20: The combined effect of network complexity and amount of evidence
on the correlation between subjective and normative probability.
Conclusions The results offer some limited support for the causal network
model. Some programmers were found to update their beliefs normatively;
however, others did not. In addition, the degree of correspondence declined as
the complexity of the inference increased.
Normative reasoning was more likely on simple problems, and less likely
when the causal network was large, or when the participants had to integrate
multiple pieces of evidence. With a larger network, there will tend to be more
links to traverse to form an inference. Similarly, when multiple pieces of evidence
are provided, the decision-maker must reason along multiple paths in order to
update the probabilities. In both cases, the number of computations would
increase, which may results in less accurate subjective judgments.
Research on human problem solving consistently shows that decision-makers
have limited memory and perform limited search of the problem space (See
[Simon, 1955].). In complex problems, rather than applying normative decision rules, it seems people may rely on heuristics (See [Kahneman et al, 1982].).
The use of heuristic information processing is more likely when the problem
becomes too complex to handle efficiently using normative methods. Therefore, while normative models may provide a good description of human reasoning with simple problems (e.g. as in the discounting studies described in
[Morris and Larrick, 1995]), normative reasoning in complex problems may require computational resources beyond the capacity of humans. Consistent with
this view, research on discounting has shown that normative reasoning occurs
EXERCISES
179
only when the participants are able to focus on the judgment task, and that
participants insufficiently discount for alternate causes when forced to perform
multiple tasks simultaneously (See [Gilbert, 1988].).
Considerable variance in the degree of correspondence was also observed
across participants, suggesting that individual differences may play a role in the
use of Bayes’ Rule. Normative reasoning may be more likely among individuals
with greater working memory, more experience with the problem domain, or
certain decision-making styles. For example, individuals who are high in need
for cognition, seem more likely than others to carefully consider multiple factors
before reaching a decision (See [Petty and Cacioppo, 1986].). Future research
should investigate such factors as how working memory might moderate the
relationship (correspondence) between normative and subjective probabilities.
That is, it should investigate whether the relationship increases with the amount
of working memory.
Experience in the problem domain is possibly a key determinant of normative
reasoning. As individuals develop expertise in a domain, it seems they learn to
process information more efficiently, freeing up the cognitive resources needed for
normative reasoning (See [Ackerman, 1987].). A limitation of the current study
was that participants had only limited familiarity with the problem domain.
While all participants had experience programming, and were at least somewhat
familiar with the type of programs involved, they were not familiar with the
details of the system in which the program operated. When working on a
program of his or her own creation, a programmer will probably have a much
deeper and more easily accessible knowledge base about the potential problems.
Therefore, complex reasoning about causes and effects may be more easy to
perform, and responses may more closely match normative predictions. An
improvement for future research would be to involve the participants in the
definition of the problem.
EXERCISES
Section 3.1
Exercise 3.1 Compute P (x1|w1) assuming the Bayesian network in Figure
3.2.
Exercise 3.2 Compute P (t1|w1) assuming the Bayesian network in Figure 3.3.
Section 3.2
180
CHAPTER 3. INFERENCE: DISCRETE VARIABLES
Exercise 3.3 Relative to the proof of Theorem 3.1, show
X
P (x|z)P (x|nZ , dT ) =
z
X
z
P (x|z)
P (z|nZ )P (nZ )P (dT |z)P (z)
.
P (z)P (nZ , dT )
Exercise 3.4 Given the initialized Bayesian network in Figure 3.7 (b), use
Algorithm 3.1 to instantiate H for h1 and then C for c2.
Exercise 3.5 Prove Theorem 3.2.
Exercise 3.6 Given the initialized Bayesian network in Figure 3.10 (b), instantiate B for b1 and then A for a2.
Exercise 3.7 Given the initialized Bayesian network in Figure 3.10 (b), instantiate A for a1 and then B for b2.
Exercise 3.8 Consider Figure 3.1, which appears at the beginning of this chapter. Use the method of conditioning to compute the conditional probabilities of
all other nodes in the network when F is instantiated for f 1 and C is instantiated for c1.
Section 3.3
Exercise 3.9 Assuming the Bayesian network in Figure 3.13, compute the following:
1. P (Z = 1|X1 = 1, X2 = 2, X3 = 2, X4 = 2).
2. P (Z = 1|X1 = 2, X2 = 1, X3 = 1, X4 = 2).
3. P (Z = 1|X1 = 2, X2 = 1, X3 = 1, X4 = 1).
Exercise 3.10 Derive Formulas 3.4, 3.5, 3.6, and 3.7.
Section 3.5
Exercise 3.11 Show what was left as an exercise in Example 3.16.
Exercise 3.12 Show what was left as an exercise in Example 3.18.
Chapter 4
More Inference Algorithms
In this chapter, we further investigate algorithms for doing inference in Bayesian
networks. So far we have considered only discrete random variables. However,
as illustrated in Section 4.1, in many cases it is an idealization to assume a variable can assume only discrete values. After illustrating the use of continuous
variables in Bayesian networks, that section develops an algorithm for doing
inference with continuous variables. Recall from Section 3.6 that the problem
of inference in Bayesian networks is N P -hard. So for some networks none of
our exact inference algorithms will be efficient. In light of this, researchers have
developed approximation algorithms for inference Bayesian networks. Section
4.2 shows an approximate inference algorithm. Besides being interested in the
conditional probabilities of every variable given a set of findings, we are often
interested in the most probable explanation for the findings. The process of determining the most probable explanation for a set of findings is called abductive
inference and is discussed in Section 4.3.
4.1
Continuous Variable Inference
Suppose a medical application requires a variables that represents a patient’s
calcium level. If we felt that it takes only three ranges to model significant
differences in patients’ reactions to calcium level, we may assign the variable
three values as follows:
Value
decreased
normal
increased
Serum Calcium Level (mg/100ml)
less than 9
9 to 10.5
above 10.5
If we later realized that three values does not adequately model the situation, we
may decide on five values, seven values, or even more. Clearly, the more values
assigned to a variable the slower the processing time. At some point it would be
more prudent to simply treat the variable as having a continuous range. Next we
181
182
CHAPTER 4. MORE INFERENCE ALGORITHMS
0.3
0.2
0.1
-4
-2
0
2 x
4
Figure 4.1: The standard normal density function.
develop an inference algorithm for the case where the variables are continuous.
Before giving the algorithm, we show a simple example illustrating how inference
can be done with continuous variables. Since our algorithm manipulates normal
(Gaussian) density functions, we first review the normal distribution and give a
theorem concerning it.
4.1.1
The Normal Distribution
Recall the definition of the normal distribution:
Definition 4.1 The normal density function with parameters µ and σ,
where −∞ < µ < ∞ and σ > 0, is
−
1
ρ(x) = √
e
2πσ
(x − µ)2
2σ2
− ∞ < x < ∞,
(4.1)
and is denoted N (x; µ, σ2 ).
A random variables X that has this density function is said to have a normal
distribution.
If the random variable X has the normal density function, then
E(X) = µ
and
V (X) = σ2 .
The density function N (x; 0, 12 ) is called the standard normal density
function . Figure 4.1 shows this density function.
The following theorem states properties of the normal density function needed
to do Bayesian inference with variables that have normal distributions:
4.1. CONTINUOUS VARIABLE INFERENCE
X
kX(x) = N(x;40,52)
Y
kY(y|x) = N(y;10x,302)
183
Figure 4.2: A Bayesian network containing continous random variables.
Theorem 4.1 These equalities hold for the normal density function:
N (x; µ, σ2 ) = N (µ; x, σ2 )
µ
¶
µ σ2
1
N x; , 2
a
a a
µ
¶
σ22 µ1 + σ21 µ2 σ 21 σ22
2
2
, 2
N (x; µ1 , σ1 )N (x; µ2 , σ2 ) = kN x;
σ21 + σ22
σ1 + σ22
N (ax; µ, σ2 ) =
where k does not depend on x.
Z
N (x; µ1 , σ 21 )N (x; y, σ22 )dx = N(y; µ1 , σ21 + σ22 ).
(4.2)
(4.3)
(4.4)
(4.5)
x
Proof. The proof is left as an exercise.
4.1.2
An Example Concerning Continuous Variables
Next we present an example of Bayesian inference with continuous random
variables.
Example 4.1 Suppose you are considering taking a job that pays $10 an hour
and you expect to work 40 hours per week. However, you are not guaranteed 40
hours, and you estimate the number of hours actually worked in a week to be
normally distributed with mean 40 and standard deviation 5. You have not yet
fully investigated the benefits such as bonus pay and nontaxable deductions such
as contributions to a retirement program, etc. However, you estimate these other
influences on your gross taxable weekly income to also be normally distributed
with mean 0 (That is, you feel they about offset.) and standard deviation 30.
184
CHAPTER 4. MORE INFERENCE ALGORITHMS
Furthermore, you assume that these other influences are independent of your
hours worked.
First let’s determine your expected gross taxable weekly income and its standard deviation. The number of hours worked X is normally distributed with
density function ρX (x) = N (x; 40, 52 ), the other influences W on your pay are
normally distributed with density function ρW (w) = N (w; 0, 302 ), and X and
W are independent. Your gross taxable weekly income Y is given by
y = w + 10x.
Let ρY (y|x) denote the conditional density function of Y given X = x. The
results just obtained imply ρY (y|x) is normally distributed with expected value
and variance as follows:
E(Y |x) = E(W |x) + 10x
= E(W ) + 10x
= 0 + 10x = 10x
and
V (Y |x) = V (W |x)
= V (W )
= 302 .
The second equality in both cases is due to the fact that X and W are independent. We have shown that ρY (y|x) = N (y; 10x, 302 ). The Bayesian network in
Figure 4.2 summarizes these results. Note that W is not shown in the network.
Rather W is represented implicitly in the probabilistic relationship between X
and Y . Were it not for W , Y would be a deterministic function of X. We
compute the density function ρY (y) for your weekly income from the values in
that network as follows:
Z
ρY (y|x)ρX (x)dx
ρY (y) =
x
=
Z
N (y; 10x, 302 )N (x; 40, 52 )dx
x
=
Z
N (10x; y, 302 )N (x; 40, 52 )dx
x
=
=
=
=
µ
¶
y 302
N x; , 2 N(x; 40, 52 )dx
10 10
x
µ
¶
y
302
1
N
; 40, 52 + 2
10
10
10
µ
·
¸¶
302
10
N y; (10) (40) , 102 52 + 2
10
10
N(y; 400, 3400).
1
10
Z
4.1. CONTINUOUS VARIABLE INFERENCE
185
The 3rd through 6th equalities above are due to Equalities 4.2, 4.3, 4.5, and 4.3
respectively. We conclude that the expected value
√ of your gross taxable weekly
income is $400 and the standard deviation is 3400 ≈ 58.
Example 4.2 Suppose next that your first check turns out to be for $300, and
this seems low to you. That is, you don’t recall exactly how many hours you
worked, but you feel that it should have been enough to make your income exceed $300. To investigate the matter, you can determine the distribution of
your weekly hours given that the income has this value, and decide whether this
distribution seems reasonable. Towards that end, we have
ρX (x|Y = 300) =
=
=
=
=
=
=
ρY (300|x)ρX (x)
ρY (300)
N (300; 10x, 302 )N (x; 40, 52 )
ρY (300)
N (10x; 300, 302 )N (x; 40, 52 )
ρY (300)
µ
¶
300 302
1
N x;
, 2 N (x; 40, 52 )
10
10 10
ρY (300)
¡
¢
1
N x; 30, 32 N (x; 40, 52 )
10
ρY (300)
µ
¶
k
52 30 + 32 40 32 52
N x;
, 2
10ρY (300)
32 + 52
3 + 52
N (x; 32.65, 6.62) .
The 3rd equality is due to Equality 4.2, the 4th is due to Equality 4.3, the 6th
is due to Equality 4.4, and the last is due to the fact that ρX (x|Y = 300) and
N (x; 32.65, 6.62) are both density functions, which means their integrals over x
must both equal 1, and therefore 102 ρ k(300) = 1. So the expected value of the
Y
√
number of hours you worked is 32.65 and the standard deviation is 6.62 ≈ 2.57.
4.1.3
An Algorithm for Continuous Variables
We will show an algorithm for inference with continuous variables in singlyconnected Bayesian networks in which the value of each variable is a linear
function of the values of its parents. That is, if PAX is the set of parents of X,
then
X
bXZ z,
(4.6)
x = wX +
Z∈PAX
where WX has density function N (w; 0, σ 2WX ), and WX is independent of each
Z. The variable WX represents the uncertainty in X’s value given values of
X’s parents. For each root X, we specify its density function N (x; µX , σ2X ).
186
CHAPTER 4. MORE INFERENCE ALGORITHMS
A density function equal to N (x; µX , 0) means we know the root’s value, while
a density function equal to N (x; 0, ∞) means complete uncertainty as to the
root’s value. Note that σ2WX is the variance of X conditional on values of its
parents. So the conditional density function of X is
X
bXZ z, σ2WX ).
ρ(x|paX ) = N(x,
Z∈PAX
When an infinite variance is used in an expression, we take the limit of
the expression containing the infinite variance. For example, if σ2 = ∞ and σ2
appears in an expression, we take the limit as σ 2 approaches ∞ of the expression.
Examples of this appear after we give the algorithm. All infinite variances
represent the same limit. That is, if we specify N (x; 0, ∞) and N (y; 0, ∞), in
both cases ∞ represents a variable t in an expression for which we take the
limit as t → ∞ of the expression. The assumption is that our uncertainty as
to the value of X is exactly the same as our uncertainty as to the value of
Y . Given this, if we wanted to represent a large but not infinite variance for
both variables, we would not use a variance of say 1, 000, 000 to represent our
uncertainty as to the value of X and a variance of ln(1, 000, 000) to represent
our uncertainty as to the value of Y . Rather we would use 1, 000, 000 in both
cases. In the same way, our limits are assumed to be the same. Of course if it
better models the problem, the calculations could be done using different limits,
and we would sometimes get different results.
A Bayesian network of the type just described is called a Gaussian Bayesian
network. The linear relationship (Equality 4.6) used in Gaussian Bayesian networks has been used in causal models in economics [Joereskog, 1982], in structural equations in psychology [Bentler, 1980], and in path analysis in sociology
and genetics [Kenny, 1979], [Wright, 1921].
Before giving the algorithm, we show the formulas used in the it. To avoid
clutter, in the following formulas we use σ to represent a variance rather than
a standard deviation.
The formula for X is as follows:
x = wX +
X
bXZ z
Z∈PAX
The λ and π values for X are as follows:
"
#−1
X
1
λ
σX =
σλU X
U ∈CH
X
µλX = σλX
X µλ
UX
λ
σ
UX
U ∈CH
X
σπX = σWX +
X
Z∈PAX
b2XZ σπXZ
4.1. CONTINUOUS VARIABLE INFERENCE
µπX =
X
187
bXZ µπXZ .
Z∈PAX
The variance and expectation for X are as follows:
σπX σ λX
π
σX + σλX
σX =
µX =
σ π µλX + σ λX µπX
.
σ πX + σλX
The π messages Z sends to a child X is as follows:
σπXZ

1
= π
σZ
µπXZ =
X
+
Y ∈CHZ −{X}
µπZ
σπZ
+
1
σπZ
+
−1
1 
σλY Z
X
Y ∈CHZ −{X}
X
Y ∈CHZ −{X}
µλY Z
σ λY Z
1
.
σ λY Z
The λ messages X sends to a parent Y are as follows:


X
1
b2Y Z σπY Z 
σλY X = 2 σλY + σWY +
bY X
Z∈PAY −{X}
µλY X =

1  λ
µY
bY X
−
X
Z∈PAY −{X}
When V is instantiated for v̂, we set

bY Z µπY Z  .
σπV = σλV = σV = 0
µπV = µλV = µV = v̂.
Next we present the algorithm. You are asked to prove it is correct in
Exercise 4.2. The proof proceeds similar to that in Section 3.2.1, and can be
found in [Pearl, 1988].
Algorithm 4.1 Inference With Continuous Variables
Problem: Given a singly-connected Bayesian network containing continuous
variables, determine the expected value and variance of each node conditional on specified values of nodes in some subset.
188
CHAPTER 4. MORE INFERENCE ALGORITHMS
Inputs: Singly-connected Bayesian network (G, P ) containing continuous variables, where G = (V, E), and a set of values a of a subset A ⊆ V.
Outputs: The Bayesian network (G, P ) updated according to the values in a.
All expectations and variances, including those in messages, are considered
part of the network.
void initial_net (Bayesian-network& (G, P ) where G = (V, E),
set-of-variables& A,
set-of-variable-values& a)
{
A = ∅; a = ∅;
for (each X ∈ V) {
σλX = ∞; µλX = 0;
// Compute λ values.
for (each parent Z of X)
σλXZ = ∞; µλXZ = 0;
// Do nothing if X is a root.
// Compute λ messages.
for (each child Y of X)
σπY X = ∞; µπY X = 0;
// Initialize π messages.
}
for (each root R) {
σR|a = σR ; µR|a = µR ;
// Compute variance and
// expectation for R.
σπR = σ R ; µπR = µR ;
// Compute R’s π values.
for (each child X of R)
send-π-msg(R, X);
}
}
void update_tree (Bayesian-network& (G, P ) where G = (V, E),
set-of-variables& A,
set-of-variable-values& a,
variable V , variable-value v̂)
4.1. CONTINUOUS VARIABLE INFERENCE
189
{
A = A ∪ {V }; a = a∪{v̂};
// Add V to A.
σπV = 0; σ λV = 0; σV |a = 0;
// Instantiate V for v̂.
µπV = v̂; µλV = v̂; µV |a = v̂;
for (each parent Z of V such that Z ∈
/ A)
send-λ-msg(V, Z);
for (each child X of V )
send-π-msg(V, X);
}
void send_λ_msg(node Y , node X)
{
h
i
P
σλY X = b21 σ λY + σWY + Z∈PAY −{X} b2Y Z σπY Z ;
// For simplicity (G, P )
// is not shown as input.
// Y sends X a λ message.
YX
µλY X =
σλX =
1
bY X
hP
1
U ∈CHX σλ
UX
µλX = σλX
σ X|a =
h
i
P
µλY − Z∈PAY −{X} bY Z µπY Z ;
P
i−1
;
µλ
UX
;
U ∈CHX σλ
UX
λ
σπ
X σX
λ ;
σπ
X +σ X
µX|a =
λ
λ π
σπ
X µX +σ X µX
;
λ
σπ
+σ
X
X
for (each parent Z of X such that Z ∈
/ A)
send_λ_msg(X, Z);
for (each child W of X such that W 6= Y )
send_π_msg(X, W );
}
// Compute X’s λ values.
// Compute variance
// and
// expectation for X.
190
CHAPTER 4. MORE INFERENCE ALGORITHMS
void send_π_message(node Z, node X)
{
i−1
h
P
σπXZ = σ1π + Y ∈CHZ −{X} σλ1
;
Z
µπXZ =
µπ
Z
σπ
Z
+
1
σπ
Z
+
P
Y ∈CHZ −{X}
P
Y ∈CHZ −{X}
P
// Z sends X a π message.
YZ
µλ
YZ
σλ
YZ
1
σλ
YZ
;
if (X ∈
/ A) {
P
σπX = σWX + Z∈PAX b2XZ σπXZ ;
µπX =
// For simplicity (G, P )
// is not shown as input.
Z∈PAX
// Compute X’s π values.
bXZ µπXZ ;
σX|a =
λ
σπ
X σX
λ ;
σπ
+σ
X
X
µX|a =
λ
λ π
σπ
X µX +σ X µX
;
λ
σπ
X +σ X
// Compute variance
// and
// expectation for X.
for (each child Y of X)
send_π_msg(X, Y );
}
if not (σX = ∞)
for (each parent W of X
such that W 6= Z and W ∈
/ A)
send_λ_msg(X, W );
//
//
//
//
Do not send λ messages
to X’s other parents if X
and all of X’s descendents
are uninstantiated.
}
As mentioned previously, the calculations with ∞ in Algorithm 4.1 are done
by taking limits, and every specified infinity represents the same variable approaching ∞. For example, if σ πP = ∞, µλP = 8000, σλP = ∞, and µπP = 0,
then
σπP µλP + σλP µπP
σπP + σλP
t × 8000 + t × 0
t+t
1 × 8000 + 1 × 0
= lim
t→∞
1+1
8000
= 4000.
= lim
t→∞ 2
=
lim
t→∞
As mentioned previously,we could let different infinite variances represent different limits, and thereby possibly get different results. For example, we could
4.1. CONTINUOUS VARIABLE INFERENCE
191
replace σπP by t and σλP by ln(t). If we did this, we would obtain
σπP µλP + σλP µπP
σπP + σλP
=
=
lim
t × 8000 + ln(t) × 0
t + ln(t)
lim
1 × 8000 +
t→∞
t→∞
1+
ln(t)
t
ln(t)
t
×0
= 8000.
Henceforth, our specified infinite variances always represent the same limit.
Since λ and π messages and values are used in other computations, we assign
variables values that are multiplies of infinity when it is indicated. For example,
if
σλDP = 0 + 3002 + ∞ + ∞,
we would make 2∞ the value of σ λDP so that 2t would be used in an expression
containing σλDP .
Next we show examples of applying Algorithm 4.1.
Example 4.3 We will redo the determinations in Example 4.1 using Algorithm
4.1 rather than directly as done in that example. Figure 4.3 (a) shows the same
network as Figure 4.2; however, it explicitly shows the parameters specified for a
Gaussian Bayesian network. The values of the parameters in Figure 4.2, which
are the ones in the general specification of a Bayesian network, can be obtained
from the parameters in Figure 4.3 (a). Indeed, we did that in Example 4.1. In
general, we show Gaussian Bayesian networks as in Figure 4.3 (a).
First we show the steps when the network is initialized.
The call
initial_tree((G, P ), A, a);
results in the following steps:
192
CHAPTER 4. MORE INFERENCE ALGORITHMS
A = ∅;
a = ∅;
σλX = ∞; µλX = 0;
// Compute λ values.
σλY = ∞; µλY = 0
σλY X = ∞; µλY X = 0;
// Compute λ messages.
σπY X = ∞; µπY X = 0;
// Compute π messages.
σX|a = 52 ; µX|a = 40;
// Compute µX |a and σX |a.
σπX = 52 ; µπR = 40;
// Compute X’s π values.
send_π_msg(X, Y );
Example 4.4 The call
send_π_msg(X, Y );
results in the following steps:
σπY X =
µπY X =
h
1
σπ
X
µπ
X
σπ
X
1
σπ
X
i−1
= σπX = 52 ;
// X sends Y a π message.
= µπX = 40;
σπY = σWY + b2Y X σπY X
// Compute Y ’s π value.
= 302 + 102 × 52 = 3400 = 58.312 ;
µπY = bY X µπY X = 10 × 40 = 400;
σY |a =
λ
σπ
Y σY
λ
σπ
Y +σ Y
µY |a =
λ
λ π
σπ
Y µY +σ Y µ Y
λ
σπ
+σ
Y
Y
3400×t
t→∞ 3400+t
= lim
= 3400;
// Compute variance
// and expectation for Y .
3400×0+t×400
3400+t
t→∞
= lim
= 400;
The initialized network is shown in Figure 4.3 (b). Note that we obtained the
same result as in Example 4.1.
Next we instantiate Y for 300 in the network in Figure 4.3 (b).
4.1. CONTINUOUS VARIABLE INFERENCE
FX = 52
:X = 40
X
X
FWY= 302
F8YX = 4
8
YX = 0
FBYX = 52
B
YX = 40
Y
F8X = 4
:8X = 0
FBX = 52
:BX = 40
FX|a = 52
:X|a = 40
9:
bYX = 10
Y
193
8:
FY|a = 58.312 FBY = 58.312
:Y|a = 400
:BY = 400
(a)
F8Y = 4
:8Y = 0
(b)
X
FX|a = 2.572
:X|a = 32.65
FBX = 52
:BX = 40
FBYX = 52
B
= 40
YX
F8YX = 32
8
= 30
YX
9:
Y
F8X = 32
:8X = 30
8:
FBY = 0
:BY = 300
FY|a = 0
:Y|a = 300
F8Y = 0
:8Y = 300
(c)
Figure 4.3: A Bayesian network modeling the relationship between hours work
and taxable income is in (a), the initialized network is in (b), and the network
after Y is instantiated for 300 is in (c).
194
CHAPTER 4. MORE INFERENCE ALGORITHMS
The call
update_tree((G, P ), A, a, Y, 300);
results in the following steps:
A = ∅ ∪ {Y } = {Y };
a = ∅ ∪ {300} = {300};
σπY = σλY = σY |a = 0;
// Instantiate Y for 300.
µπY = µλY = µY |a = 300;
send_λ_msg(Y, X);
The call
send_λ_msg(Y, X);
results in the following steps:
σλY X =
1
b2Y X
µλY X =
1
bY X
h
σλ
YX
σλX =
1
£ λ
¤
σY + σ WY =
£
¤
µλY =
i−1
1
10
1
100
[0 + 900] = 9;
// Y sends X a λ
// message.
[300] = 30;
= 9;
// Compute X’s λ
// values.
µλ
µλX = σλX σ Yλ X = 9 30
9 = 30;
YX
σX|a =
λ
σπ
X σX
λ
σπ
X +σ X
µX|a =
λ
λ π
σπ
X µX +σ X µX
λ
σπ
+σ
X
X
=
25×9
25+9
=
= 6.62 = 2.572 ;
25×30+9×40
25+9
= 32.65;
// Compute variance
// and expectation
// for X.
The updated network is shown in Figure 4.3 (c). Note that we obtained the same
result as in Example 4.2.
Example 4.5 This example is based on an example in [Pearl, 1988]. Suppose
we have the following random variables:
Variable
P
D
What the Variable Represents
Wholesale price
Dealer’s asking price
4.1. CONTINUOUS VARIABLE INFERENCE
FX = 4
:X = 0
P
P
FBDP = 4
B
=0
DP
FWD = 3002
D
F8DP = 4
8
=0
DP
8:
FD|a = 4
:D|a = 0
(a)
F8P = 4
:8P = 0
FBP = 4
:BP = 0
9:
bDP = 1
D
FP|a = 4
:P|a = 0
195
FBD = 4
:BD = 0
F8D = 4
:8D = 0
FBP = 4
:BP = 0
F8P = 3002
:8P = 8000
(b)
P
FP|a = 3002
:P|a = 8000
FBDP = 4
B
=0
DP
F8DP = 3002
8
DP = 8000
9:
D
8:
FD|a = 0
FBD = 0
:D|a = 8000 :BD = 8000
F8Y = 0
:8Y = 8000
(c)
Figure 4.4: The Bayesian network in (a) models the relationship between a car
dealer’s asking price for a given vehicle and the wholesale price of the vehicle. The network in (b) is after initialization, and the one in (c) is after D is
instantiated for $8, 000.
We are modeling the relationship between a car dealer’s asking price for a given
vehicle and the wholesale price of the vehicle. We assume
d = wD + p
σD = 3002
where WD is distributed N (wD ; 0, σWD ). The idea is that in past years, the
dealer has based its asking price on the mean profit from the last year, but there
has been variation, and this variation is represented by the variables WD . The
Bayesian network representing this model appears in Figure 4.4 (a). Figure 4.4
(b) shows the network after initialization. We show the result of learning that
the asking price is $8, 000.
The call
196
CHAPTER 4. MORE INFERENCE ALGORITHMS
update_tree((G, P ), A, a, D, 8000);
results in the following steps:
A = ∅ ∪ {D} = {D};
a = ∅ ∪ {8000} = {8000};
σπD = σλD = σD|a = 0;
// Instantiate D for 8000.
µπD = µλD = µD|a = 8000;
send_λ_msg(D, P );
The call
send_λ_msg(D, P );
results in the following steps:
σλDP =
1
b2DP
µλDP =
1
bDP
h
σλ
DP
σλP =
1
µλ
£ λ
¤
σD + σ WD =
£ λ¤
µD =
i−1
1
1
1
1
£
¤
0 + 3002 = 3002 ;
// D sends P a λ
// message.
[8000] = 8000;
= 3002 ;
// Compute P ’s λ
// values.
8000
= 3002 300
µλP = σλP σDP
2 = 8000;
λ
DP
σP |a =
λ
σπ
P σP
λ
σπ
P +σ P
µP |a =
λ
λ π
σπ
P µ P +σ P µP
λ
σπ
+σ
P
P
t×3002
2
t→∞ t+300
= lim
= 3002 ;
// Compute variance
// and expectation
=
2
×0
lim t×8000+300
2
t+300
t→∞
= 8000;
// for P .
The updated network is shown in Figure 4.4 (c). Note that the expected
value of P is the value of D, and the variance of P is the variance owing to the
variability W .
Example 4.6 Suppose we have the following random variables:
Variable
P
M
D
What the Variable Represents
Wholesale price
Mean profit per car realized by Dealer in past year
Dealer’s asking price
4.1. CONTINUOUS VARIABLE INFERENCE
FM = 4
:M = 0
197
FP = 4
:P = 0
M
P
bDM = 1
bDP = 1
D
FWD = 3002
(a)
FM|a = 4/2
:M|a = 4000
FBM = 4
:BM = 0
FP|a = 4/2
: P|a = 4000
F8M = 4
:8M = 8000
M
9
FBP = 4
:B P = 0
F8P = 4
:8P = 8000
P
FBDP = 4
B
DP = 0
8
DP = 4
8
DP = 8000
FBDM = 4
:BDM = 0
9:
F
8:
F8DM = 4
8
DM = 8000
8:
D
FD|a = 0
: D|a = 8000
FBD = 0
: BD = 8000
F8D = 0
: 8D = 8000
(b)
FM|a = 0
F8M = 0
FBM = 0
: M|a = 1000 :BM = 1000 : 8M = 1000
FP|a = 3002
:P|a = 7000
M
FBDM
:BDM
9
FBP = 4
:BP = 0
F8P = 3002
:8P = 7000
P
FBDP = 4
B
DP = 0
9:
8:F
=0
= 1000
F8DM = 4
8
DM = 8000
8
DP
8:
8
DP
= 3002
= 7000
D
FD|a = 0
: D|a = 8000
FBD = 0
: BD = 8000
F8D = 0
: 8D = 8000
(c)
Figure 4.5: The Bayesian network in (a) models the relationship between a car
dealer’s asking price for a given vehicle, the wholesale price of the vehicle, and
the dealer’s mean profit in the past year. The network in (b) is after initialization
and after D is instantiated for $8, 000, and the network in (c) is after M is also
instantiated for $1000.
198
CHAPTER 4. MORE INFERENCE ALGORITHMS
We are now modeling the situation where the car dealer’s asking price for a
given vehicle is based both on the wholesale price of the vehicle and the mean
profit per car realized by the dealer in the past year. We assume
d = wD + p + m
σ D = 3002
where WD is distributed N (wD ; 0, σWD ). The Bayesian network representing
this model appears in Figure 4.5 (a). We do not show the initialized network
since its appearance should now be apparent. We show the result of learning that
the asking price is $8, 000.
The call
update_tree((G, P ), A, a, D, 8000);
results in the following steps:
A = ∅ ∪ {D} = {D};
a = ∅ ∪ {8000} = {8000};
σπD = σλD = σD|a = 0;
µπD = µλD = µD|a = 8000;
send_λ_msg(D, P );
send_λ_msg(D, M );
The call
send_λ_msg(D, P );
results in the following steps:
// Instantiate D for 8000.
4.1. CONTINUOUS VARIABLE INFERENCE
σ λDP =
1
b2DP
£ λ
¤
σD + σWD + b2DM σπDM
µλDP =
=
[8000 − 1 × 0] = 8000;
i−1
σ λP =
h
µλP
µλ
σ λP σ DP
λ
DP
=
// λ message.
£ λ
¤
µD − bDM µπDM
1
bDP
1
1
// D sends P a
£
¤
0 + 3002 + 1 × t = ∞;
1
t→∞ 1
= lim
199
1
σλ
DP
= lim
t→∞
=
£
£ 1 ¤−1
t
lim t 8000
t
t→∞
σ P |a =
λ
σπ
P σP
λ
σπ
P +σ P
µP |a =
λ
λ π
σπ
P µP +σ P µP
λ
σπ
+σ
P
P
t×t
t→∞ t+t
= lim
¤
= ∞;
// Compute P ’s
// λ values.
= 8000;
t
t→∞ 2
= lim
=
∞
;
2
// Compute variance
// and expectation
= lim
t→∞
// for P .
t×8000+t×0
t+t
=
8000
2
= 4000;
Clearly, the call send_λ_msg(D, M ) results in the same values for M as
we just calculated for P .
The updated network is shown in Figure 4.5 (b). Note that the expected
values of P and M are both 4000, which is half the value of D. Note further
that each variable still has infinite variance owing to uncertainty as to the value
of the other variable.
Notice in the previous example that D has two parents, and each of their
expected values is half of the value of D. What would happen if D had a third
parent F , bDF = 1, and F also had an infinite prior variance? In this case,
σλDP
=
=
1 £
b2DP
σλD + σWD + b2DM σπDM + b2DF σπDF
¤
1£
0 + 3002 + 1 × t + 1 × t = 2∞.
t→∞ 1
lim
¤
This means σλP also equals 2∞, and therefore,
µP |a
σ πP µλP + σλP µπP
σπP + σλP
8000
t × 8000 + 2t × 0
=
= 2667.
= lim
t→∞
t + 2t
3
=
It is not hard to see that if there are k parents of D, all bDX ’s are 1 and all
prior variances are infinite, and we instantiate D for d, then the expected value
of each parent is d/k.
200
CHAPTER 4. MORE INFERENCE ALGORITHMS
Example 4.7 Next we instantiate M for 1000 in the network in Figure 4.5 (b).
The call
update_tree((G, P ), A, a, M, 1000);
results in the following steps:
A = {D} ∪ {M} = {D, M };
a = {8000} ∪ {1000} = {8000, 1000};
σπM = σλM = σ M|a = 0;
// Instantiate M for
// 1000.
µπM = µλM = µM |a = 1000;
send_π_msg(M, D);
The call
send_π_msg(M, D);
results in the following steps:
σπDM =
µπDM
=
h
1
σπ
M
µπ
M
σπ
M
1
σπ
M
i−1
= σ πM = 0;
= µπM = 1000;
send_λ_msg(D, P );
The call
send_λ_msg(D, P );
results in the following steps:
// M sends D a π message.
4.1. CONTINUOUS VARIABLE INFERENCE
σ λDP =
1
b2DP
µλDP =
1
bDP
h
σλ
DP
σ λP =
1
µλP = σ λP
h
£ λ
¤
σD + σWD + b2DP σπDM =
£ λ
¤
µD − bDMm µπDM =
+
µλ
DP
σλ
DP
1
σλ
EP
+
i−1
µλ
EP
σλ
EP
σ P |a =
λ
σπ
P σP
λ
σπ
+σ
P
P
µP |a =
λ
λ π
σπ
P µP +σ P µP
λ
σπ
P +σ P
£
1
1
1
2
t→∞ 300
= lim
i
= 3002 lim
1
1
t×3002
2
t→∞ t+300
= lim
t→∞
£
¤
0 + 3002 + 0 = 3002 ;
[8000 − 1000] = 7000;
+
¤
1 −1
t
£ 7000
2
t→∞ 300
= lim
201
= 3002 ;
t×7000+3002 ×0
t+3002
+
= 3002 ;
10000
t
¤
= 7000;
= 7000;
The final network is shown in Figure 4.5 (c). Note that the expected value
of P is the difference between the value of D and the value of M . Note further
that the variance of P is now simply the variance of WD .
Example 4.8 Suppose we have the following random variables:
Variable
P
D
E
What the Variable Represents
Wholesale price
Dealer-1’s asking price
Dealer-2’s asking price
We are now modeling the situation where there are two dealers, and for each
the asking price is based only on the wholesale price and not on the mean profit
realized in the past year. We assume
d = wD + p
e = wE + p
σD = 3002
σ E = 10002 ,
where WD is distributed N(wD ; 0, σWD ) and WE is distributed N(wE ; 0, σWE ).
The Bayesian network representing this model appears in Figure 4.6 (a). Figure
4.6 (b) shows the network after we learn the asking prices of Dealer-1 and Dealer2 in the past year are $8, 000 and $10, 000 respectively. We do not show the
calculations of the message values in that network because these calculations are
just like those in Example 4.5. We only show the computations done when P
receives both its λ messages. They are as follows:
202
CHAPTER 4. MORE INFERENCE ALGORITHMS
FP = 4
:P = 0
P
bDP = 1
bEP = 1
D
E
FWE = 1000 2
FWD = 3002
(a)
FP|a = 2872
:P|a = 8145
FBP = 4 F 8P = 2872
8
:BP = 0 : P = 8145
P
FBDP = 4
:BDP = 0
9
F
8:
8
=
DP
8
=
DP
FBEP = 4
B
=0
EP
9:
F 8EP = 1000 2
8
EP = 10000
3002
8000
8:
D
E
FBD = 0
:BD = 8000
FD|a = 0
:D|a = 8000
F8D = 0
:8D = 8000
FE|a = 0
: E|a = 10000
FBE = 0
:BE = 10000
F8E = 0
: 8E = 10000
(b)
Figure 4.6: The Bayesian network in (a) models the relationship between two
car dealers’ asking price for a given vehicle and the wholesale price of the vehicle.
The network in (b) is after initialization and D and E are instantiated for $8, 000
and $10, 000 respectively.
σλP =
h
1
σλ
DP
µλP = σλP
h
+
1
σλ
EP
µλ
DP
σλ
DP
+
i−1
µλ
EP
σλ
EP
σP |a =
λ
σπ
P σP
λ
σπ
+σ
P
P
µP |a =
λ
λ π
σπ
P µ P +σ P µP
λ
σπ
+σ
P
P
=
i
£
1
3002
= 3002
t×2872
2
t→∞ t+297
= lim
= lim
t→∞
+
¤−1
1
10002
£ 8000
3002
+
= 2872 ;
10000
10002
= 2872 ;
t×8145+2872 ×0
t+2872
¤
= 8145;
= 8145;
Notice the expected value of the wholesale price is closer to the asking price
of the dealer with less variability.
4.1. CONTINUOUS VARIABLE INFERENCE
FM = 4
:M = 0
203
FP = 4
:P = 0
M
FN = 4
:N = 0
P
bDP = 1
bDM = 1
N
bEP = 1
bEN = 1
D
E
FWD = 300 2
FWE = 1000 2
(a)
FM|a = 4/2
:M|a = 4000
FBM = 4
:BM = 0
FBDM = 4
B
=0
DM
9:
8
F
8:
DM
8
DM
F8M = 4
: 8M = 8000
M
=4
= 8000
FP|a = 4/3
: P|a = 6000
FBDP = 4
B
=0
DP
F8DP = 4
: 8DP = 8000
9:
FBP = 4
:BP = 0
P
F8P = 4/2
:8P = 9000
FBEP = 4
B
=0
EP
9:
8:F
8
8
EP
8
EP
D
FD|a = 0
: D|a = 8000
FBD = 0
: BD = 8000
FN|a = 4/2
: N|a = 5000
N
=4
= 10000
F BN = 4
:BN = 0
F8N = 4
: 8N = 10000
FBEN = 4
B
EN = 0
8
=4
EN
8
= 10000
EN
9:
F
8:
E
F 8D = 0
: 8D = 8000
FE|a = 0
: E|a = 10000
FBE =0
: BE = 10000
F 8E = 0
: 8E = 10000
(b)
FM|a = 0
F8M = 0
FBM = 0
: M|a = 1000 : BM = 1000 :8M = 1000
9:
FBDM = 0
= 1000
DM
B
M
F8DM = 4
8
DM = 8000
FP|a = 287 2
:P|a = 7165
FBDP = 4
B
=0
DP
9:
FBP = 4
:BP = 0
P
F8DP = 300 2
8
DP = 7000
8
D
FD|a = 0
: D|a = 8000
FBD = 0
: BD = 8000
FBEP = 4
B
EP = 0
9:
F
8:
8
8:
8:
FN|a = 0
: N|a = 1000
F8P = 287 2
:8P = 7165
EP
EP
N
= 1000 2
= 9000
F8N = 0
F BN = 0
8
: BN = 1000 : N = 1000
FBEN = 0
B
= 1000
EN
9:
F
8:
8
8
EN
EN
=4
= 10000
E
F 8D = 0
: 8D = 8000
FE|a = 0
: E|a = 10000
F BE = 0
: BE = 10000
F 8E = 0
: 8E = 10000
(c)
Figure 4.7: The Bayesian network in (a) models the relationship between two car
dealers’ asking price for a given vehicle, the wholesale price of the vehicle, and
the mean profit per car realized by each dealer in the past year. The network in
(b) is after initialization and D and E are instantiated for $8, 000 and $10, 000
respectively, and the one in (c) is after M and N are also instantiated for $1, 000.
204
CHAPTER 4. MORE INFERENCE ALGORITHMS
Example 4.9 Suppose we have the following random variables:
Variable
P
M
D
N
E
What the Variable Represents
Wholesale price
Mean profit per car realized by Dealer-1 in past year
Dealer-1’s asking price
Mean profit per car realized by Dealer-2 in past year
Dealer-2’s asking price
We are now modeling the situation where we have two dealer’s, and for each
the asking price is based both on the wholesale price and the mean profit per car
realized by the dealer in the past year. We assume
σ D = 3002
d = wD + p + m
σ E = 10002 ,
e = wE + p + n
where WD is distributed N (wD ; 0, σWD ) and WE is distributed N (wE ; 0, σWE ).
The Bayesian network representing this model appears in Figure 4.7 (a). Figure
4.7 (b) shows the network after initialization and after we learn the asking prices
of Dealer-1 and Dealer-2 in the past year are $8, 000 and $10, 000 respectively.
For that network, we only show the computations when P receives its λ messages
because all other computations are exactly like those in Example 4.6. They are
as follows:
σλP =
h
1
σλ
DP
µλP = σλP
h
+
1
σλ
EP
µλ
DP
σλ
DP
i−1
µλ
EP
σλ
EP
+
σP |a =
λ
σπ
P σP
λ
σπ
+σ
P
P
µP |a =
λ
λ π
σπ
P µ P +σ P µP
λ
σπ
P +σ P
= lim
t→∞
i
£1
t
t
t→∞ 2
= lim
t× 2t
t
t+
t→∞
2
= lim
= lim
t→∞
=
∞
3 ;
+
¤
1 −1
t
£ 8000
t
+
t×9000+ 2t ×0
t+ 2t
=
∞
2 ;
10000
t
¤
= 9000;
= 6000;
Note in the previous example that the expected value of the wholesale price
is greater than half of the asking price of either dealer. What would happen of
D had a third parent F , bDF = 1, and F also had an infinite prior variance? In
this case,
σλDP
=
=
So
σ λP =
·
1
σλDP
¤
1 £ λ
σD + σWD + b2DM σπDM + b2DF σπDF
b2DP
¤
1£
0 + 3002 + 1 × t + 1 × t = 2∞.
t→∞ 1
lim
+
1
σλEP
¸−1
= lim
t→∞
·
1
1
+
2t
t
¸−1
=
2∞
3
4.2. APPROXIMATE INFERENCE
µλP = σλP
and
µP |a =
·
µλ
µλDP
+ EP
λ
σDP
σλEP
¸
205
·
¸
2t 8000 10000
+
= 9333,
t→∞ 3
2t
t
= lim
t × 9333 + 2t
σ πP µλP + σλP µπP
3 ×0
=
lim
= 5600.
2t
π
λ
t→∞
σP + σP
t+ 3
Notice that the expected value of the wholesale price has decreased. It is not
hard to see that, as the number of such parents of D approaches infinity, the
expected value of the wholesale price approaches half the value of E.
Example 4.10 Next we instantiate both M and N for 1000 in the network in
Figure 4.7 (b). The resultant network appears in Figure 4.7 (c). It is left as an
exercise to obtain that network.
4.2
Approximate Inference
As mentioned at the beginning of this chapter, since the problem of inference
in Bayesian networks is N P -hard researchers have developed approximation
algorithms for inference in Bayesian networks. One way to do approximate
inference is by sampling data items, using a pseudorandom number generator,
according to the probability distribution in the network, and then approximate
the conditional probabilities of interest using this sample. This method is called
stochastic simulation. We discuss this method here. Another method is to
use deterministic search, which generates the sample systematically. You are
referred to [Castillo et al, 1997] for a discussion of that method.
First we review sampling. After that we show a basic sampling algorithm
for Bayesian networks called logic sampling. Finally, we improve the basic
algorithm.
4.2.1
A Brief Review of Sampling
We can learn something about probabilities from data when the probabilities
are relative frequencies, which were discussed briefly in Section 1.1.1. The following two examples illustrate the difference between relative frequencies and
probabilities that are not relative frequencies.
Example 4.11 Suppose the Chicago Bulls are about to play in the 7th game of
the NBA finals, and I assess the probability that they will win to be .6. I also
feel there is a .9 probability there will be a big crowd celebrating at my favorite
restaurant that night if they do win. However, even if they lose, I feel there
might be a big crowd because a lot of people may show up to lick their wounds.
So I assign a probability of .3 to a big crowd if they lose. I can represent
this probability distribution with the two-node Bayesian network in Figure 4.8.
Suppose I work all day, drive straight to my restaurant without finding out the
result of the game, and see a big crowd overflowing into the parking lot. I can
then use Bayes’ Theorem to compute my conditional probability they did indeed
win. It is left as an exercise to do so.
206
CHAPTER 4. MORE INFERENCE ALGORITHMS
Bulls
Crowd
P(Bulls = win) = .6
P(Crowd = big|Bulls = win) = .9
P(Crowd = big|Bulls = lose) = .3
Figure 4.8: A Bayesian network in which the probabilities cannot be learned
from data.
Example 4.12 Recall Example 1.23 in which we discussed the following situation: Joe had a routine diagnostic chest X-ray required of all new employees at
Colonial Bank, and the X-ray came back positive for lung cancer. The test had
a true positive rate of .6 and a false positive rate of .02. That is,
P (T est = positive|LungCancer = present) = .6
P (T est = positive|LungCancer = absent) = .02.
Furthermore, the only information about Joe, before he took the test, was that
he was one of a class of employees who took the test routinely required of new
employees. So, when he learned only 1 out of every 1000 new employees has lung
cancer, he assigned about .001 to P (LungCancer = present). He then employed
Bayes’ theorem to compute P (LungCancer = present|T est = positive). Recall
in Example 1.30 we represented this probability distribution with the two-node
Bayesian network in Figure 1.8. It is shown again in Figure 4.9.
There are fundamental differences in the probabilities in the previous two
examples. In Example 4.12, we have experiments we can repeat, which have
distinct outcomes, and our knowledge about the conditions of each experiment
is the same every time it is executed. Richard von Mises was the first to formalize
this notion of repeated identical experiments. He said [von Mises, 1928]
The term is ‘the collective’, and it denotes a sequence of uniform
events or processes which differ by certain observable attributes, say
colours, numbers, or anything else. [p. 12]
I, not von Mises, put the word ‘collective’ in bold face above. The classical
example of a collective is an infinite sequence of tosses of the same coin. Each
time we toss the coin, our knowledge about the conditions of the toss is the
same (assuming we do not sometimes ‘cheat’ by, for example, holding it close
4.2. APPROXIMATE INFERENCE
207
Lung
P(LungCancer = present) = .001
Cancer
P(Test = positive|LungCancer = present) = .6
Test
P(Test = positive|LungCancer = absent) = .02
Figure 4.9: A Bayesian network in which the probabilities can be learned from
data.
to the ground and trying to flip it just once). Of course, something is different
in the tosses (e.g. the distance from the ground, the torque we put on the
coin, etc.) because otherwise the coin would always land heads or always land
tails. But we are not aware of these differences. Our knowledge concerning the
conditions of the experiment is always the same. Von Mises argued that, in
such repeated experiments, the relative frequency of each outcome approaches
a limit and he called that limit the probability of the outcome. As mentioned
in Section 1.1.1, in 1946 J.E. Kerrich conducted many experiments indicating
the relative frequency does indeed appear to approach a limit.
Note that the collective (infinite sequence) only exists in theory. We never
will toss the coin indefinitely. Rather the theory assumes there is a propensity
for the coin to land heads, and, as the number of tosses approaches infinity, the
fraction of heads approaches that propensity. For example, if m is the number
of times we toss the coin, Sm is the number of heads, and p is the true value of
P ({heads}),
Sm
.
(4.7)
p = lim
m→∞ m
Note further that a collective is only defined relative to a random process,
which, in the von Mises theory, is defined to be a repeatable experiment for
which the infinite sequence of outcomes is assumed to be a random sequence.
Intuitively, a random sequence is one which shows no regularity or pattern.
For example, the finite binary sequence ‘1011101100’ appears random, whereas
the sequence ‘1010101010’ does not because it has the pattern ‘10’ repeated five
times. There is evidence that experiments like coins tossing and dice throwing
are indeed random processes. Namely, in 1971 G.R. Iversen et al ran many
experiments with dice indicating the sequence of outcomes is random. It is
believed that unbiased sampling also yields a random sequence and is therefore
a random process. See [van Lambalgen, M., 1987] for a thorough discussion of
this matter, including a formal definition of random sequence. Neapolitan [1990]
208
CHAPTER 4. MORE INFERENCE ALGORITHMS
provides a more intuitive, less mathematical treatment. We close here with an
example of a nonrandom process. I prefer to exercise at my health club on
Tuesday, Thursday, and Saturday. However, if I miss a day, I usually make up
for it the following day. If we track the days I exercise, we will find a pattern
because the process is not random.
Under the assumption that the relative frequency approaches a limit and
that a random sequence is generated, in 1928 R. von Mises was able to derive
the rules of probability theory and the result that the trials are probabilistically
independent. In terms of relative frequencies, what does it mean for the trials
to be independent? The following example illustrates what it means. Suppose
we develop sequences of length 20 (or any other number) by repeatedly tossing
a coin 20 times. Then we separate the set of all these sequences into disjoint
subsets such that the sequences in each subset all have the same outcome on
the first 19 tosses. Independence means the relative frequency of heads on the
20th toss is the same in all the subsets (in the limit).
Let’s discuss the probabilities in Examples 4.11 and 4.12 relative to the concept of a collective. In Example 4.12, we have three collectives. First, we have
the collective consisting of an infinite sequence of individuals who apply for a
job at Colonial Bank, where the observable attribute is whether lung cancer is
present. Next we have the collective consisting of an infinite sequence of individuals who both apply for a job at Colonial Bank and have lung cancer, where
the observable attribute is whether a chest X-ray is positive. Finally, we have
the collective consisting of an infinite sequence of individuals who both apply
for a job at Colonial Bank and do not have lung cancer, where the observable
attribute is again whether a chest X-ray is positive. According to the von Mises
theory, in each case there is propensity for a given outcome to occur and the
relative frequency of that outcome will approach that propensity. Sampling
techniques estimate this propensity from a finite set of observations. In accordance with standard statistical practice, we use the term random sample(or
simply sample) to denote the set of observations. In a mathematically rigorous
treatment of sampling (as we do in Chapter 6), ‘sample’ is also used to denote
the set of random variables whose values are the finite set of observations. We
will use the term both ways, and it will be clear from the context which we
mean. To distinguish propensities from subjective probabilities, we often use
the term relative frequency rather than the term probability to refer to a
propensity.
In the case of Example 4.11 (the Bulls game), I certainly base my probabilities on previous observations, namely how well the Bulls have played in the
past, how big crowds were at my restaurant after other big games, etc. But we
do not have collectives. We cannot repeat this particular Bulls’ game with our
knowledge about its outcome the same. So sampling techniques are not directly
relevant to learning probabilities like those in the DAG in Figure 4.8. If we did
obtain data on crowds in my restaurant on evenings of similar Bulls’ games, we
could possibly roughly apply the techniques but this might prove to be complex.
We sometimes call a collective a population. Before leaving this topic, we
note the difference between a collective and a finite population. There are
4.2. APPROXIMATE INFERENCE
209
currently a finite number of smokers in the world. The fraction of them with lung
cancer is the probability (in the sense of a ratio) of a current smoker having lung
cancer. The propensity (relative frequency) of a smoker having lung cancer may
not be exactly equal to this ratio. Rather the ratio is just an estimate of that
propensity. When doing statistical inference, we sometimes want to estimate the
ratio in a finite population from a sample of the population, and other times
we want to estimate a propensity from a finite sequence of observations. For
example, TV raters ordinarily want to estimate the actual fraction of people in
a nation watching a show from a sample of those people. On the other hand,
medical scientists want to estimate the propensity with which smokers have lung
cancer from a finite sequence of smokers. One can create a collective from a finite
population by returning a sampled item back to the population before sampling
the next item. This is called ‘sampling with replacement’. In practice it
is rarely done, but ordinarily the finite population is so large that statisticians
make the simplifying assumption it is done. That is, they do not replace the
item, but still assume the ratio is unchanged for the next item sampled. In this
text, we are always concerned with propensities rather than current ratios. So
this simplifying assumption does not concern us.
Estimating a relative frequency from a sample seems straightforward. That
is, we simply use Sm /m as our estimate, where m is the number of trials and
Sm is the number of successes. However, there is a problem in determining our
confidence in the estimate. That is, the von Mises theory only says the limit in
Expression 4.7 physically exists and is p. It is not a mathematical limit in that,
given an ² > 0, it offers no means for finding an M (²) such that
¯
¯
¯
¯
¯p − Sm ¯ < ²
for m > M (²).
¯
m¯
Mathematical probability theory enables us to determine confidence in our
estimate of p. First, if we assume the trials are probabilistically independent,
we can prove that Sm /m is the maximum likelihood (ML) value of p. That
is, if d is a set of results of m trials, and P (d : p̂) denotes the probability of d if
the probability of success were p̂, then Sm /m is the value of p̂ that maximizes
P (d : p̂). Furthermore, we can prove the weak and strong laws of large numbers.
The weak law says the following. Given ², δ > 0
¯
¶
µ¯
¯
Sm ¯¯
2
¯
<² >1−δ
for m > 2 .
P ¯p −
¯
m
δ²
So mathematically we have a means of finding an M(², δ).
The weak law is not applied directly to obtain confidence in our estimate.
Rather we obtain a confidence interval using the following result, which is obtained in a standard statistics text such as [Brownlee, 1965]. Suppose we have
m independent trials, the probability of success on each trial is p, and we have
k successes. Let
0<β<1
α = (1 − β)/2
210
CHAPTER 4. MORE INFERENCE ALGORITHMS
θ1 =
θ2 =
kFα (2k, 2[m − k + 1])
m − k + 1 + kFα (2x, 2[m − k + 1])
k
,
(m − k + 1)F1−α (2[m − k + 1], 2k) + k
where F is the F distribution. Then
(θ1 , θ 2 )
is a β % confidence interval for p.
This means β % of the time the interval generated will contain p.
Example 4.13 Recall Example 1.2 in which we discussed repeatedly tossing a
thumbtack. Suppose we toss it 30 times and it lands heads (i..e. on its head) 8
times. It is left as an exercise to derive the following 95% confidence interval
for p, the probability of heads:
(.123, .459)
Since 95% of the time we will obtain an interval that contains p, we are pretty
confident p is in this interval.
One should not conclude that mathematical probability theory somehow
proves Sm /m will be close to p, and that therefore we have no need for the von
Mises theory. Without some assumption about Sm /m approaching p, the mathematical result would say nothing about what is happening in the world. For
example, without some such assumption, our explanation of confidence intervals
would become the following: Suppose we have a sample space determined by
m identically distributed independent discrete random variables, where p is the
probability each of them assumes its first value. Consider the random variable
whose possible values are the probability intervals obtained using the method for
calculating a β % confidence interval. Then β is the probability that the value of
this random variable is an interval containing p. This result says nothing about
what will happen when, for example, we toss a thumbtack m times. However, if
we assume that the probability of an event is the limit of the relative frequency
with which the event occurs in the world, this means that if we repeatedly did
the experiment of tossing the thumbtack m times, in the limit 95% of the time
we will generate an interval containing p, which is how we described confidence
intervals above.
Some probabilists find fault with the von Mises theory because it assumes
the relative frequency definitely approaches p. For example [Ash, 1970] [p. 2]
says
...an attempt at a frequency definition of probability will cause
trouble. If Sn is the number of occurrences of an event in n independent performances of an experiment, we expect physically that
the relative frequency Sn /n should converge to a limit; however, we
cannot assert that the limit exists in a mathematical sense. In the
4.2. APPROXIMATE INFERENCE
211
case of the tossing of an unbiased coin, we expect Sn /n → 1/2, but a
conceivable outcome of the process is that the coin will keep coming
up heads forever. In other words, it is possible that Sn /n → 1, or
that Sn /n → any number between 0 and 1, or that Sn /n has no
limit at all.
As mentioned previously, in 1946 J.E. Kerrich conducted many experiments
using games of chance indicating that the relative frequency does appear to
approach a limit. However, if it is only most likely this would happen, any
such experiment may indicate that it does. So to resolve the objection posed by
Ash, in 1992 R.E. Neapolitan obtained von Mises’ results concerning the rules
of probability by assuming Sm /m → p only in the sense of the weak law of large
numbers.
In Chapter 6, we further discuss estimating relative frequencies by sampling.
In that chapter, we use an approach which incorporates one’s prior belief concerning a relative frequency into the computation of the estimate obtained from
the sample. That approach is called Bayesian, whereas the approach presented
here is often called frequentist.
4.2.2
Logic Sampling
Suppose p is the fraction of black balls in an urn containing white and black
ball. We could sample with replacement from the urn, and, according to the
theory discussed in the previous section, use k/m as an estimate of p, where
of m balls sampled k are black. Alternatively, assuming we have a function
random that returns a pseudorandom number between 0 and 1 according to
the uniform distribution, we could write the following computer simulation of
sampling m balls:
k = 0;
for (i = 1; i <= m; i + +)
if random() < p
k + +;
p̂ = k/m;
Our estimate of p is p̂. This is called a simulation because we are not sampling
from an actual distribution, but rather are using pseudorandom numbers to
imitate the process of sampling. Of course, the previous simulation has no
value because, if we knew p, we would have no need to estimate it. The purpose
of discussing this simulation is for illustration.
Suppose next that we have the Bayesian network in Figure 4.10, and we
wish to compute P (y1). Instead of computing it directly, we could do the
following simulation to estimate it. First we determine all probabilities in the
212
CHAPTER 4. MORE INFERENCE ALGORITHMS
X
P(x1) = .7
P(y1| x1) = .8
Y
P(y1|x2) = .4
Figure 4.10: A Bayesian Network
joint distribution as follows:
P (x1, y1)
P (x1, y2)
P (x2, y1)
P (x2, y2)
=
=
=
=
P (y1|x1)P (x1) = (.2)(.5) = .1
P (y2|x1)P (x1) = (.8)(.5) = .4
P (y1|x2)P (x2) = (.6)(.5) = .3
P (y2|x2)P (x2). = (.4)(.5) = .2.
Next we use our pseudorandom number generator to obtain m values of (x, y)
according to this distribution. If we let k be the number of tuples containing
y1, we then have the following estimate:
P̂ (y1) =
k
.
m
Example 4.14 Suppose we have the Bayesian network in Figure 4.10, m = 7,
and we generate the data in the following table:
Case
1
2
3
4
5
6
7
X
x2
x1
x1
x2
x1
x2
x2
Y
y1
y1
y2
y1
y2
y1
y2
Since y1 occurs in 4 cases, our estimate is
P̂ (y1) =
4
k
= .
m
7
4.2. APPROXIMATE INFERENCE
213
In a large network, we could never compute all the values in the joint distribution. So instead of using the method just described, we could obtain each
case by first generating a value x̃ of X using P (x), and then generating a value
ỹ of Y using P (y|x̃). For each tuple (x̃, ỹ), x̃ will occur P (x̃) fraction of the
time in the limit; of those occurrences, ỹ will occur P (ỹ|x̃) fraction of the time
in the limit. So (x̃, ỹ) will occur P (ỹ|x̃)P (x̃) = P (x̃, ỹ) fraction of the time in
the limit, which is what we want. The following is a high-level algorithm for
this method:
k = 0;
for (i = 1; i <= m; i + +) {
generate a value x̃ of X using P (x);
generate a value ỹ of Y using P (y|x̃);
if (ỹ == y1)
k + +;
}
P̂ (y1) = k/m;
Example 4.15 Suppose we have the Bayesian network in Figure 4.10, and m =
3. The following shows one possible outcome of our simulation:
1. We generate a value of X using P (x1) = .5. Suppose we find x1 occurs.
We then generate a value for Y using P (y1|x1) = .2. Suppose we find y2
occurs. We do not increment k.
2. We generate a value for X using P (x1) = .5. Suppose we find x2 occurs.
We then generate a value for Y using P (y1|x2) = .6. Suppose we find y1
occurs. We increment k to 1.
3. We generate a value for X using P (x1) = .5. Suppose we find x1 occurs.
We then generate a value for Y using P (y1|x1) = .2. Suppose we find y1
occurs. We increment k to 2.
Our final estimate is
P̂ (y1) = 2/3.
We can use the method just presented to also estimate P (x1|y1). However,
we must throw out any cases that have Y = y2. The following high level
algorithm does this:
k = 0;
for (i = 1; i <= m; i + +) {
repeat
generate a value x̃ of X using P (x1);
generate a value ỹ of Y using P (y1|x̃);
until (ỹ == y1);
if (x̃ == x1)
k + +;
}
P̂ (x1|y1) = k/m;
214
CHAPTER 4. MORE INFERENCE ALGORITHMS
Example 4.16 Suppose we have the Bayesian network in Figure 4.10, m = 5,
and the preceding algorithm generates the data in the following table:
Case
1
2
3
4
5
6
7
X
x2
x1
x1
x2
x1
x2
x2
Y
y1
y1
y2
y1
y1
y1
y2
Cases 3 and 7 are rejected because Y = y2. So k = 2, and we have
P̂ (x1|y1) =
2
k
= .
m
5
You may wonder why we regenerate a new value for X when we generate a
value of y2 for Y . That is, you may ask why can’t we just retain our old value,
and keep generating a value of Y until we get y1. If we did this, in the limit
we would simply get X generated according to its prior probability because the
X values we kept have nothing to do with the Y values we generate. Recall we
want to generate X values according to P (x1|y1). The following table shows
values our algorithm would generate if the first 10 generated cases corresponded
exactly to the distribution in Figure 4.10.
Case
1
2
3
4
5
6
7
8
9
10
X
x1
x1
x1
x1
x1
x2
x2
x2
x2
x2
Y
y1
y2
y2
y2
y2
y1
y1
y1
y2
y2
Our algorithm will reject Cases 2, 3, 4, 5, 9, and 10, and it will estimate P (x1|y1)
to be 1/4 (using a value of 4 for m), which you can check is the correct values.
However, if we simply kept all X values and each time kept generating Y values
until we got y1, our estimate would be 5/10 (using a value of 10 for m).
The method just outlined is easily extended to an algorithm for doing approximate inference in Bayesian networks. The algorithm first orders the nodes
according to an ancestral ordering, which you should recall is an ordering of
the nodes such that,if Z is a descendent of Y , then Z follows Y in the ordering.
The algorithm, called logic sampling, was developed in [Henrion, 1988].
4.2. APPROXIMATE INFERENCE
215
Algorithm 4.2 Approximate Inference Using Logic Sampling
Problem: Given a Bayesian network, determine the probabilities of the values
of each node conditional on specified values of the nodes in some subset.
Inputs: Bayesian network (G, P ), where G = (V, E), and a set of values a of
a subset A ⊆ V.
Outputs: Estimates of the conditional probabilities of the values of each node
in V − A given A = a.
void logic_sampling (Bayesian-network& (G, P ) where G = (V, E),
set-of-variables A,
set-of-variable-values a,
estimates& P̂ (xj |a))
{
order the n nodes in V in an ancestral ordering;
for (each Xj ∈ V − A)
// xjk is
for (k = 1; k <= # of values in Xj ’s space; k + +)
// the kth
set # of occurrences of xjk to 0;
for (i = 1; i <= m; i + +) {
// value in
j = 1;
// Xj ’s space.
while (j <= n) {
generate a value x̃j for Xj using
P (xj |p̃aj ) where p̃aj is the
set of values generated for Xj ’s parents;
if (Xj ∈ A && x̃j 6= the value of Xj ∈ a)
j = 1;
else
j + +;
}
for (each Xj ∈ V − A)
for (k = 1; k <= # of values in Xj ’s space; k + +)
if (xjk == x̃j )
add 1 to # of occurrences of xjk ;
}
for (each Xj ∈ V − A)
for (k = 1; k <= # of values in Xj ’s space; k + +)
# of occurrences of xjk
P̂ (xjk |a) =
;
m
}
216
CHAPTER 4. MORE INFERENCE ALGORITHMS
P(X1 = 1) = .5
X1
P(X2 = 1|X1 = 1) = .8
X2
P(X2 = 1|X1 = 2) = .1
P(X3 = 1|X1 = 1) = .7
X3
X4
P(X3 = 1|X1 = 2) = .4
X5
P(X4 = 1|X2 = 1) = .6
P(X5 = 1|X2 = 1, X3 = 1) = .1
P(X4 = 1|X2 = 2) = .1
P(X5 = 1|X2 = 1, X3 = 2) = .2
P(X5 = 1|X2 = 2, X3 = 1) = .3
P(X5 = 1|X2 = 2, X3 = 2) = .4
Figure 4.11: The Bayeisan network discussed in Example 4.17.
Example 4.17 Suppose we have the Bayesian network in Figure 4.11. Note
that we are using 1 and 2 as the values of all variables. If we instantiate X3
to 1 and X4 to 2, then A = {X3 , X4 } and a = {1, 2}. An application of the
preceding algorithm with m = 4 may yield the following data:
Case
1
2
3
4
5
6
7
X1
1
1
1
2
2
2
1
X2
2
2
2
1
2
1
1
X3
1
2
1
1
1
2
1
X4
2
X5
2
2
1
2
1
2
1
2
Note that Case 2 never obtained values for X4 or X5 because the value generated
for X3 was not its value in a. We had similar results for Cases 4 and 6. The
resultant estimates are as follows:
P̂ (X1 = 1|X3 = 1, X4 = 2) =
3
4
P̂ (X2 = 1|X3 = 1, X4 = 2) =
1
4
4.2. APPROXIMATE INFERENCE
217
(b)
(a)
Figure 4.12: We can estimate the probability of picking a black ball from the
urn in (a) by sampling from the urn in (b).
P̂ (X5 = 1|X3 = 1, X4 = 2) =
4.2.3
2
.
4
Likelihood Weighting
A problem with logic sampling is that we need to reject cases which do not have
the evidence variables A instantiated for a. We reject each case with probability
1 − P (a). So if the probability of the evidence is low, we will reject many cases.
Next we present a method called likelihood weighting, which circumvents
this problem. Before showing how the method is used in Bayesian networks, we
present a simple example illustrating the method.
Consider the two urns in Figure 4.12. Let Color be a random variable whose
value is black if we pick a black ball and whose value is white if we pick a white
ball. Furthermore, let P (black) be the probability Color = black for the urn in
Figure 4.12 (a), and let P 0 (black) be the probability Color = black for the urn
in Figure 4.12 (b). We will show how to estimate P (black) by sampling with
replacement from the urn in Figure 4.12 (b). Suppose we create a sample of size
m by sampling from that urn, and k of the balls are black. Instead of adding 1
each time we sample a black ball, we add score(black), where
score(black) =
P (black)
.
P 0 (black)
By adding score(black) each time we sample a black ball and then dividing the
result by m, we obtain
k × score(black)
.
m
218
CHAPTER 4. MORE INFERENCE ALGORITHMS
In the limit, we have
k × score(black)
lim
m→∞
m
k
=
³
P (black)
P 0 (black)
´
lim
m→∞
m
µ
¶
P (black)
k
=
lim
P 0 (black) m→∞ m
µ
¶
P (black)
=
P 0 (black) = P (black),
P 0 (black)
which is what we want. Therefore, if we use [k × score(black)] /m as an estimate
of P (black), we will have convergence. However, for any finite sample, we would
not necessarily have
k × score(black) (m − k) × score(white)
+
= 1.
m
m
So instead we simply determine k × score(black) and (m − k) × score(white),
and then normalize to obtain our estimate. The following example illustrates
this.
Example 4.18 Suppose we sample 100 balls with replacement from the urn in
Figure 4.12 (b) and 72 are black. We have
score(black) =
2/3
P (black)
=
= 8/9
P 0 (black)
3/4
score(white) =
1/3
P (white)
=
= 4/3.
P 0 (white)
1/4
k × score(black) = 72(8/9) = 64
(m − k) × score(white) = 28(4/3) =
112
,
3
So our estimate of P (black) is given me
P̂ (black) =
64
= .632.
64 + 112/3
The previous sampling strategy for estimating P (black) has no practical
value because we had to know P (black) to do the estimation. However, next we
show how the method can be applied to Bayesian networks in a case where we
do not know the probabilities of interest.
Let (G, P ), where G = (V, E), be a Bayesian network, V = {X1 , X2 , . . . Xn },
A ⊆ V, W = V − A, a and w be sets of values of the variables in A and W
respectively, and v = w ∪ a. Then
P (w, a)
P (a)
= αP (v)
= αP (xn |pan )P (xn−1 |pan−1 ) · · · P (x2 |pa2 )P (x1 |pa1 ),
P (w|a) =
4.2. APPROXIMATE INFERENCE
219
where xi and pai are respectively the values of Xi and its parents PAi in v,
and α is a normative constant. Now suppose we let P 0 (w) be the probability
distribution obtained by taking the product of only the conditional distributions
in the Bayesian network of the variables in W with the evidence variables A
clamped to a. That is,
Y
P (xi |pai ),
P 0 (w) =
Xi ∈W
where again the values of all variables are their values in v. We now define
score(w) ≡
=
=
P (w|a)
αP 0 (w)
αP (xn |pan )P (xn−1 |pan−1 ) · · · P (x2 |pa2 )P (x1 |pa1 )
Q
α Xi ∈W P (xi |pai )
Y
P (xi |pai ).
Xi ∈A
Notice that we have eliminated the normalizing constant α in the score. Since
we eventually normalize to obtain our probability estimates from the scores,
there is no reason to include the constant. Before giving an algorithm for this
method, we show an example.
Example 4.19 Suppose we have the Bayesian network in Figure 4.11. To estimate P (Xj = xj |X3 = 1, X4 = 2) for j = 1, 2, and 5 using the method
just described, we first clamp X3 to 1 and X4 to 2. Then we generate values
of the other variables according to the distributions in the network. For example, the first case is generated as follows: We initially generate a value of
X1 according to P (X1 = 1) = .5. Let’s say we get a value of 2 for X1 . We
then generate a value of X2 according to P (X2 = 1|X1 = 2) = .1. Let’s say
we get a value of 2 for X2 . Finally, we generate a value of X5 according to
P (X5 = 1|X2 = 2, X3 = 1) = .3. Note that the value of X2 is 2 because this is
the value that was generated, while the value of X3 is 1 because X3 is clamped
to this value. Let’s say we get a value of 1 for X5 . The score score0 of a case
is defined to be the score of the value of w for that case. For example, the score
of the first case (just discussed) is given by
score0 (Case 1) = score(X1 = 2, X2 = 2, X5 = 1)
= P (X4 = 2|X2 = 2)P (X3 = 1|X1 = 2)
= (.9)(.4) = .36.
The following table shows possible data for the first 4 cases and the corresponding
scores:
Case
1
2
3
4
X1
2
1
2
1
X2
2
1
1
1
X3
1
1
1
1
X4
2
2
2
2
X5
1
2
2
1
score0
.36
.28
.16
.28
220
CHAPTER 4. MORE INFERENCE ALGORITHMS
Finally we estimate the conditional probability of the value of any particular
variable by normalizing the sum of the scores for the cases containing that value.
For example,
P̂ (X1 = 1|X3 = 1, X4 = 2) ∝ [score0 (Case 2) + score0 (Case 4)]
∝ [.28 + .28] = .56
P̂ (X1 = 2|X3 = 1, X4 = 2) ∝ [score0 (Case 1) + score0 (Case 3)]
∝ [.36 + .16] = .52.
So
P̂ (X1 = 1|X3 = 1, X4 = 2) =
.56
= .52.
.56 + .52
It is left as an exercise to do the computations that estimate the conditional
probabilities of the other variables.
Next we give an algorithm for the likelihood weighing method.
Algorithm 4.3 Approximate Inference Using Likelihood Weighting
Problem: Given a Bayesian network, determine the probabilities of the values
of each node conditional on specified values of the nodes in some subset.
Inputs: Bayesian network (G, P ), where G = (V, E), and a set of values a of
a subset A ⊆ V.
Outputs: Estimates of the conditional probabilities of the values of each node
in V − A given A = a.
void like_weight (Bayesian-network& (G, P ) where G = (V, E),
set-of-variables A,
set-of-variable-values a,
estimates& P̂ (xj |a))
4.3. ABDUCTIVE INFERENCE
221
{
}
order the n nodes in V in an ancestral order;
for (each Xj ∈ V − A)
for (k = 1; k <= # of values in Xj ’s space; k + +)
P̂ (xjk |a) = 0;
for (each Xj ∈ A)
set x̃j to the value of Xj in a;
for (i = 1; i <= m; i + +) {
for (j = 1; j <= n; j + +) {
/ A)
if (Xj ∈
generate a value x̃j for Xj using
P (xjk |p̃aj ) where p̃aj is the
set of values generated for Xj ’s parents;
}
Q
score = Xj ∈A P (x̃j |p̃aj );
for (each Xj ∈ V − A);
for (k = 1; k <= # of values in Xj ’s space; k + +)
if (xjk == x̃j )
P̂ (xjk |a) = P̂ (xjk |a) + score;
}
for (each Xj ∈ V − A)
for (k = 1; k <= # of values in Xj ’s space; k + +)
normalize P̂ (xjk |a);
//
//
//
//
xjk is
the kth
value in
Xj ’s space.
// Use all
// values of k.
Algorithm 4.3 was developed independently in [Fung and Chang, 1990] and
[Shachter and Peot, 1990]. It is proven in [Dagum and Luby, 1993] that the
problem of approximate inference in Bayesian networks is N P -hard. However,
there are restricted classes of Bayesian networks which are provably amenable to
a polynomial-time solution (See [Dagum and Chavez, 1993].). Indeed, a variant
of the likelihood weighting algorithm, which is worst-case polynomial time as
long as the network does not contain extreme conditional probabilities, was
developed in [Pradham and Dagum, 1996].
4.3
Abductive Inference
First we describe abductive inference in Bayesian networks; then we present an
algorithm for it.
4.3.1
Abductive Inference in Bayesian Networks
Recall the Bayesian network discussed in Example 1.32. That network is shown
again in Figure 4.13. Recall further that the variables represent the following:
222
CHAPTER 4. MORE INFERENCE ALGORITHMS
P(h1) = .2
H
P(b1|h1) = .25
P(b1|h2) = .05
B
L
F
P(f1|b1,l1) = .75
P(f1|b1,l2) = .10
P(f1|b2,l1) = .5
P(f1|b2,l2) = .05
P(l1|h1) = .003
P(l1|h2) = .00005
C
P(c1|l1) = .6
P(c1|l2) = .02
Figure 4.13: A Bayesian nework.
Variable
H
B
L
F
C
Value
h1
h2
b1
b2
l1
l2
f1
f2
c1
c2
When the Variable Takes this Value
Patient has a smoking history
Patient does not have a smoking history
Patient has bronchitis
Patient does not have bronchitis
Patient has lung cancer
Patient does not have lung cancer
Patient is fatigued
Patient is not fatigued
Patient has a positive chest X-ray
Patient has a negative chest X-ray
We discussed this network again in the beginning of chapter 3. We noted that if a
patient had a smoking history and a positive chest X-ray, we would be interested
in the probability of that patient having lung cancer (i.e. P (l1|h1, c1)) and
having bronchitis (i.e. P (b1|h1, c1)). We went on to develop algorithms that
perform this type of inference. Besides being interested in these conditional
probabilities, a physician would be interested in the most probable explanation
for the symptoms. That is, the physician would be interested in whether it
is most probable that the patient has both lung cancer and bronchitis, has
lung cancer and does not have bronchitis, does not have lung cancer and has
bronchitis, or does not have either lung cancer or bronchitis. In general, the
physician is interested in the most probable set of diseases given some symptoms.
Similarly, in the case of an electronic circuit, we would be interested in the most
probable explanation for a failure at some point in the circuit. Another example
is the determination for the most probable cause of failure of an automobile to
4.3. ABDUCTIVE INFERENCE
223
function properly. This process of determining the most probable explanation
for a set of findings is called abductive inference. We have the following
definition specific to Bayesian networks:
Definition 4.2 Let (G, P ) where G = (V, E) be a Bayesian network, let M ⊆ V,
D ⊆ V, and M ∩ D = ∅. M is called the manifestation set and D is called the
explanation set. Let m be a set of values of the variables in M. Then a set of
values of the variables in D that maximizes
P (d|m)
is called a most probable explanation (MPE) for m. The process of determining such a set is called abductive inference.
Example 4.20 Suppose we have the Bayesian network in Figure 4.13, M =
{H, C}, D = {B, L} and m = {h1, c1}. Then a most probable explanation for
m contains values of B and H that maximize
P (bi, lj|h1, c1).
The chain rule gives us a straightforward algorithm for determining a most
probable explanation. That is, if D = {D1 , D2 , ...Dk }, M = {M1 , M2 , ...Mj }, m =
{m1 , m2 , ...mj }, and d = {d1 , d2 , ...dk } is a set of values of the variables in D,
then
P (d|m) = P (d1 , d2 , d3, ...dk |m1 , m2 , ...mj )
= P (d1 |d2 , ...dk , m1 , m2 , ...mj )P (d2 |d3 , ...dk , m1 , m2 , ...mj )
· · · P (dk |m1 , m2 , ...mj ).
We can compute all the probabilities, in the expression on the right in the
equality above, using our algorithms for doing inference in Bayesian networks.
So to determine a most probable explanation, we simply use this method to
compute the conditional probabilities of all the explanations, and then we take
the maximum.
Example 4.21 To compute a most probable explanation for the instance in
Example 4.20, we need compute the following 4 conditional probabilities:
P (b1, l1|h1, c1) = P (b1|l1, h1, c1)P (l1|h1, c1)
P (b1, l2|h1, c1) = P (b1|l2, h1, c1)P (l2|h1, c1)
P (b2, l1|h1, c1) = P (b2|l1, h1, c1)P (l1|h1, c1)
P (b2, l2|h1, c1) = P (b2|l2, h1, c1)P (l2|h1, c1).
After doing this, we take the maximum to determine a most probable explanation.
224
CHAPTER 4. MORE INFERENCE ALGORITHMS
The problem with the simple algorithm just described is that it has exponential time complexity. For example, if each variable has only two values and
their are k variables in the explanation set, we must determine the conditional
probability of 2k explanations. [Cooper, 1990] has shown that the problem of
abductive inference in Bayesian networks is NP -hard. One way to handle optimization problems such as abductive inference is to use best-first search with
branch-and-bound pruning. For many instances this technique avoids generating most of the possible explanations and is therefore efficient. This is the
method presented here. Zhaoyu and D’Ambrosio [1993] develop an algorithm
for finding the r most probable explanations in an arbitrary Bayesian network,
which does not use search.
4.3.2
A Best-First Search Algorithm for Abductive Inference
The best-first search with branch-and-bound pruning technique is used to solve
problems in which a set of items needs to be chosen so as to maximize or
minimize some function of the items. Neapolitan and Naimipour [1998] present
a general introduction to the technique. Here we only develop an algorithm for
abductive inference using the technique.
For the sake of focus, we will use medical terminology. We assume that the
explanation set consists of k possible diseases, each of which may or may not be
present in the patient. That is,
D = {D1 , D2 , ..., Dk },
We know that the patient has a certain set of values m of certain symptoms M.
Our goal is to find the set of diseases that is most probably present. Note that
these assumptions entail that each variable in the explanation set has only two
values. Let A = Di1 , Di2 , ...Dij be a subset of D. We will denote the event that
the diseases in A are present and all other diseases are absent by
A
and by
Di1 , Di2 , ...Dij .
For example, suppose there are 4 diseases. Then
D1 , D3
represents the event
D1 = present, D2 = absent, D3 = present, D4 = absent.
We call
P (D1 , D3 |m)
4.3. ABDUCTIVE INFERENCE
225
i
i
{D1}
{D1}
{D1,D2}
{D1,D2,D3}
{D1,D2}
{D1,D3}
i
{D2}
{D1}
{D2,D3}
{D2}
{D3}
i
Figure 4.14: The state space tree for abductive inference when there are 3
possible diseases.
the conditional probability of the diseases. Note that this is not consistent
with the usual use of this terminology because ordinarily it means these diseases
are present and it is not known whether others are also. Here it entails that no
other diseases are present.
We can solve the problem of determining the most probable set of diseases
(conditional on the information that some symptoms are present) by constructing a state space tree, such that each node in the tree contains a subset of
D, as follows: The root of the tree contains the empty set, the left child of the
root contains {D1 }, and the right child of the root contains the empty set. The
left child of the root’s left child contains {D1 , D2 }, and its right child contains
{D1 }. In general, we go to the left of a node a level i to include disease Di+1
and we go to the right to not include it (Note that the root is at level 0.).
Each leaf in the state space tree represents a possible solution (that is, the set
of diseases that have been included up to the leaf). To solve the problem, we
compute the conditional probability of the set of diseases at each leaf, and then
we determine which conditional probability is largest. The tree for the case of
3 possible diseases is shown in Figure 4.14.
Our goal is to avoid generating most nodes in the tree. We can often accomplish this by determining a bounding function, which, at each node, puts
an upper bound on the conditional probabilities of the sets of diseases in all descendents of the node. As we generate nodes starting from the root, we compute
226
CHAPTER 4. MORE INFERENCE ALGORITHMS
both the conditional probability of the disease set at the node and the bound for
the node. We use the bound for two purposes. First, we prune any node whose
bound is less than the greatest conditional probability found so far. Second,
we always next expand the node with the current best bound. In this way, we
can often arrive at an optimal solution faster than we would if we visited the
nodes in some predetermined order. This technique is called best-first search
with brand-and-bound pruning. Before we can illustrate the technique, we
need to find a bounding function. The following theorem accomplishes this for
a large class of instances.
Theorem 4.2 If A and A0 are two sets of diseases such that
P (A0 ) ≤ P (A),
then
P (A0 |m) ≤
P (A)
.
P (m)
Proof. According to Bayes’ Theorem
P (A0 |m) =
≤
≤
P (m|A0 )P (A0 )
P (m)
P (m|A0 )P (A)
P (m)
P (A)
.
P (m)
The first inequality is due to the assumption in the theorem, and the second is
due to the fact that any probability is less than or equal to 1. This proves the
theorem.
For a given node, let A be the set of diseases that have been included up to
the node, and for some descendent of that node, let A0 be the set of diseases that
have been included up to that descendent. Then A ⊆ A0 . Often it is reasonable
to assume that
P (A0 ) ≤ P (A)
when
A ⊆ A0 .
The reason is that usually it is at least as probable that a patient has a set
of diseases as it is that the patient has that set plus even more diseases (Recall
that these are prior probabilities before any symptoms are observed.). If we
make this assumption, Theorem 4.2 implies that
P (A0 |m) ≤
P (A)
.
P (m)
Therefore, P (A)/P (m) is an upper bound on the conditional probability of the
set of diseases in any descendent of the node.
4.3. ABDUCTIVE INFERENCE
227
Next we show an example that uses this bound to prune branches. First
we need some additional terminology. When the bound for a given node is
no better than the value of the best solution found so far, the node is called
nonpromising. Otherwise, it is called promising.
Example 4.22 Suppose there are four diseases D1 , D2 , D3 , and D4 in our
explanation set D, and we have a set of values m of the variables in our symptom
set M. The input to this example would also include a Bayesian network that
contains the probabilistic relationships among the diseases and the symptoms.
The probabilities used in this example would be computed from this Bayesian
network using an algorithm for inference in a Bayesian network. Therefore, do
not think that there is somewhere in this text where they are computed. We are
assigning arbitrary probabilities to illustrate the best-first search algorithm.
Figure 4.15 is the pruned state space tree produced by a best-first search with
branch-and-bound pruning. Probabilities have been given arbitrary values in the
tree. The conditional probability is on the top and the bound is on the bottom
at each node. The shaded node is where the best solution is found. Nodes are
labeled according to their depth and position from the left in the tree. The steps
that produce the tree follow. The value of the variable Best is the current best
solution, while P (Best|m) is its conditional probability. Our goal is to determine
a value of Best that maximizes this conditional probability. It is also assumed
arbitrarily that
p(m) = .01.
1. Visit node (0,0) (the root).
(a) Compute its conditional probability. {∅ is the empty set. This means
no diseases are present.}
P (∅|m) = .1.
{The computation would be done by another}
{algorithm. We are assigning arbitrary values.}
(b) Set
Best = ∅
P (Best|m) = .1.
(c) Compute its prior probability and bound.
P (∅) = .9.
bound =
.9
P (∅)
=
= 90.
P (m)
.01
2. Visit node (1,1).
(a) Compute its conditional probability.
P (D1 |m) = .4.
228
CHAPTER 4. MORE INFERENCE ALGORITHMS
(0,0)
i
P( |m) = .1
bound = 90
(1,1)
(1,2)
P(D1|m) = .4
bound = .9
P( |m) = .1
bound = 90
i
(2,1)
(2,2)
(2,3)
(2,4)
P(D1,D2|m) = .1
bound = .3
P(D1|m) = .4
bound = .9
P(D2|m) = .15
bound = .5
P( |m) = .1
bound = 90
i
(3,1)
(3,2)
(3,3)
(3,4)
P(D1,D3|m)=.05
bound = .1
P(D1|m) = .4
bound = .9
P(D3|m) = .1
bound = .2
P( |m) = .1
bound = 90
i
(4,1)
(4,2)
(4,3)
(4,4)
P(D1,D4|m)=.65
bound = 0
P(D1|m) = .4
bound = 0
P(D4|m) = .6
bound = 0
P( |m) = .1
bound = 0
i
Figure 4.15: The pruned state space tree produced using best-first search with
brand-and-bound pruning in Example 4.22. At each node, the conditional probability of the diseases included up to that node is at the top, and the bound
on the conditional probability that could be obtained by expanding beyond the
node is at the bottom. The shaded node is the one at which a most probable
explanation is found.
4.3. ABDUCTIVE INFERENCE
229
(b) Since .4 is greater than P (Best|m), set
Best = {D1 }
P (Best|m) = .4.
(c) Compute its prior probability and bound.
P (D1 ) = .009
bound =
P (D1 )
.009
=
= .9.
P (m)
.01
3. Visit node (1,2).
(a) Its conditional probability is simply that of its parent, namely .1.
(b) Its prior probability and bound are simply those of its parent, namely
.9 and 90.
4. Determine promising, unexpanded node with largest bound.
(a) That node is node (1,2). We visit its children next.
5. Visit node (2,3).
(a) Compute its conditional probability.
P (D2 |m) = .15.
(b) Compute its prior probability and bound.
P (D2 ) = .005.
bound =
P (D2 )
.005
=
= .5.
P (m)
.01
6. Visit node (2,4).
(a) Its conditional probability is simply that of its parent, namely .1.
(b) Its prior probability and bound are simply those of its parent, namely
.9 and 90.
7. Determine promising, unexpanded node with largest bound.
(a) That node is node (2,4). We visit its children next.
8. Visit node (3,3).
(a) Compute its conditional probability.
P (D3 |m) = .1.
230
CHAPTER 4. MORE INFERENCE ALGORITHMS
(b) Compute its prior probability and bound.
P (D3 ) = .002.
bound =
.002
P (D3 )
=
= .2.
P (m)
.01
(c) Determine it is non-promising because its bound .2 is less than .4,
the value of P (Best|m).
9. Visit node (3,4).
(a) Its conditional probability is simply that of its parent, namely .1.
(b) Its prior probability and bound are simply those of its parent, namely
.9 and 90.
10. Determine promising, unexpanded node with the largest bound.
(a) That node is node (3,4). We visit its children next.
11. Visit node (4,3).
(a) Compute its conditional probability.
P (D4 |m) = .6.
(b) Since .6 > P (Best|m), set
Best = {D4 }
P (Best|m) = .6.
(c) Set its bound to 0 because it is a leaf in the state space tree.
(d) At this point the node (2,3) becomes non-promising because its bound
.5 less than or equal to .6, the new value of P (Best|m).
12. Visit node (4,4).
(a) Its conditional probability is simply that of its parent, namely .1.
(b) Set its bound to 0 because it is a leaf in the state space tree.
13. Determine promising, unexpanded node with largest bound.
(a) That node is node (1,1). We visit its children next.
14. Visit node (2,1).
(a) Compute its conditional probability.
P (D1 , D2 |m) = .1.
4.3. ABDUCTIVE INFERENCE
231
(b) Compute its prior probability and bound.
P (D1 , D2 ) = .003
bound =
.003
P (D1 , D2 )
=
= .3.
P (m)
.01
(c) Determine it is non-promising because its bound .3 less than or equal
to .6, the value of P (Best|m).
15. Visit node (2,2).
(a) Its conditional probability is simply that of its parent, namely .4.
(b) Its prior probability and bound are simply those of its parent, namely
.009 and .9.
16. Determine promising, unexpanded node with greatest bound.
(a) The only promising, unexpanded node is node (2,2). We visit its
children next.
17. Visit node (3,1).
(a) Compute its conditional probability.
P (D1 , D3 |m) = .05.
(b) Compute its prior probability and bound.
P (D1 , D3 ) = .001
bound =
.001
P (D1 , D3 )
=
= .1.
P (m)
.01
(c) Determine it is non-promising because its bound .1 less than or equal
to .6, the value of P (Best|m).
18. Visit node (3,2).
(a) Its conditional probability is simply that of its parent, namely .4.
(b) Its prior probability and its bound are simply those of its parent,
namely .009 and .9.
19. Determine promising, unexpanded node with largest bound.
(a) The only promising, unexpanded node is node (3,2). We visit its
children next.
20. Visit node (4,1).
232
CHAPTER 4. MORE INFERENCE ALGORITHMS
(a) Compute its conditional probability.
P (D1 , D4 |m) = .65.
(b) Since .65 is greater than P (Best|m), set
Best = {D1, D4 }
P (Best|m) = .66.
(c) Set its bound to 0 because it is a leaf in the state space tree.
21. Visit node (4,2).
(a) Its conditional probability is simply that of its parent, namely .4.
(b) Set its bound to 0 because it is a leaf in the state space tree.
22. Determine promising, unexpanded node with largest bound.
(a) There are no more promising, unexpanded nodes. We are done.
We have determined that the most probable set of diseases is {D1 , D4 } and
that P (D1 , D4 |m) = .65.
A reasonable strategy in this problem would be to initially sort the diseases
in non-decreasing order according to their prior probabilities. There is no guarantee, however, that this strategy will minimize the search time. We did not
do this in the previous example and there were 15 nodes checked. In the exercises, you will establish that if the diseases were sorted, there would be 23 nodes
checked.
We present the algorithm shortly. However, first we need discuss our implementation of best-first search. This implementation uses a priority queue. In a
priority queue, the element with the highest priority is always removed next.
In best-first search applications, the element with the highest priority is the
node with the best bound. A priority queue can be implemented as a linked
list, but more efficiently as a heap. We manipulate the priority queue P Q with
the following two procedures.
Insert(P Q, X)
is a procedure that adds X to the priority queue P Q, while
Remove(P Q, X)
is a procedure that removes the node with the best bound and assigns its value
to X. When removing a node from P Q, we have a check which determines if
the bound for the node is still better than Best. This is how we determine that
a node has become non-promising after visiting the node. For example, node
(2,3) in Figure 4.15 is promising at the time we visit it. In our implementation,
4.3. ABDUCTIVE INFERENCE
233
this is when we insert it in P Q. However, it becomes non-promising when Best
takes the value .6. In our implementation, this is before we remove it from P Q.
We learn this by comparing its bound to Best after removing it from P Q. In
this way, we avoid visiting children of a node that becomes non-promising after
it is visited. Since we need the bound for a node at insertion time, at removal
time, and to order the nodes in the priority queue, we store the bound at the
node. The declaration is as follows:
struct node
{
int level;
set-of-indices A;
float bound;
};
// the node’s level in the tree
The value of the field A is the set of indices of the diseases included up to
the node. The algorithm now follows. It has come to be known as Cooper’s
Algorithm because it was developed by Greg Cooper in [Cooper, 1984].
Algorithm 4.4 Cooper’s Best-First Search Algorithm for Abductive Inference
Problem: Determine a most probable set of diseases (explanation) given a set
of symptoms. It is assumed that if a set of diseases A is a subset of a set
of diseases A0 , then
P (A0 ) ≤ P (A).
Inputs: Positive integer n, Bayesian network (G, P ) where G = (V, E), ordered
subset D of V containing n disease variables, and set of values m of the
variables in a subset M of V.
Outputs: A set Best that contains the indices of the diseases in a most probable explanation, and a variable P best that is the probability of Best given
that M = m.
void Cooper (int n,
Bayesian-network& (G, P ) where G = (V, E),
ordered-set-of-diseases D,
set-of-symptoms M,
set-of-symptom-values m,
set-of-indices& Best, float& P best)
234
CHAPTER 4. MORE INFERENCE ALGORITHMS
{
priority-queue-of-node P Q;
node X, Y ;
X.level = 0;
X.A = ∅;
Best = ∅;
P best = P (Best|m);
X.bound = bound(X);
insert(P Q, X);
while (!empty(P Q)){
remove(P Q, X);
if (X.bound > P best){
Y.level = X.level + 1;
Y.A = X.A;
put Y.level in Y.A;
if (P (Y.A|m) > P best){
Best = Y.A;
P best = P (Y.A|m);
}
Y.bound = bound(Y );
if (Y.bound > P best)
insert(P Q, Y );
Y.A = X.A;
Y.bound = bound(Y );
if (Y.bound > P best)
insert(P Q, Y );
}
}
// Set X to the root.
// Store empty set at root.
// Remove node with best bound.
// Set Y to a child of X.
// Set Y to the child that includes
// the next disease.
// Set Y to the child that does
// not include the next disease.
}
int bound (node Y )
{
if (Y.level == n)
return 0;
else
return (P (Y.A)/P (m);
}
// A leaf is non-promising.
The notation P (A) stands for the prior probability of A, P (m) stands for the
prior probability of m, and P (A|m) stands for the conditional probability of A
given m. These values would be computed from the Bayesian network (G, P )
using an algorithm for inference in a Bayesian network.
We have written the algorithm strictly according to guidelines for writing
best-first search algorithms. An improvement is possible. First, there is no need
EXERCISES
235
to call function bound for the right child of a node. The reason is that the right
child contains the same set of diseases as the node itself, which means its bound
is the same. Therefore, the right child will be pruned only if we change P best
to a value greater than or equal to this bound at the left child. We can modify
our algorithm to prune the right child when this happens, and to expand to the
right child when it does not happen
If there is more than one best solution, Algorithm 4.4 only produces one of
them. It is straightforward to modify the algorithm to produce all the best
solutions. It is also possible to modify it to produce the r most probable
solutions, where r is any positive integer. This modification is discussed in
[Neapolitan, 1990]. Furthermore, Neapolitan [1990] analyzes the algorithm in
detail.
EXERCISES
Section 4.1
Exercise 4.1 Prove Theorem 4.1.
Exercise 4.2 Prove Algorithm 4.1 is correct.
Exercise 4.3 Obtain the network in Figure 4.7 (c).
Exercise 4.4 This exercise concerns an expanded model of the auto pricing
problem discussed in Example 4.9. Suppose we have the following random variables:
Variable
P
M
D
N
E
C
K
X
R
I
What the Variable Represents
Wholesale price
Mean profit per car realized by Dealer-1 in past year
Dealer-1’s asking price
Mean profit per car realized by Dealer-2 in past year
Dealer-2’s asking price
Production cost
Marketing cost
An expert’s estimate of the production cost
An expert’s estimate of the marketing cost
Manufacturer’s profit
Suppose further that the relationships among the variables are modeled by the
Bayesian network in Figure 4.16.
a) Initialize this network.
236
CHAPTER 4. MORE INFERENCE ALGORITHMS
FC = 4
:C = 0
FI = 4
:I = 0
FK = 4
:K = 0
C
I
K
bXC = 1
bPI = 1
bPK = 1
FM = 4
:M = 0
X
bRK = 1
FN = 4
:N = 0
M
P
TX = 502
R
N
TR = 2002
TP = 100 2
bDP = 1
bDM = 1
bEP = 1
bEN = 1
D
E
TD = 300 2
TE = 10002
(a)
Figure 4.16: A Bayesian network representing an expanded model of the auto
pricing problem.
b) Suppose we learn that the expert estimates the production cost to be $5, 000
and the Marketing cost to be $2, 000. That is, we have the following instantiations:
X = 5000
R = 2000.
Update the network based on this information.
c) Suppose next we learn that Dealer-1 has an asking price of $8, 000 and
Dealer-2 has an asking price of $8, 000. That is, we now also have the following
instantiations:
D
E
= 8000
= 10000.
Update the network based on this additional information.
Section 4.2
EXERCISES
237
Exercise 4.5 In Example 4.11, it was left as an exercise to use Bayes’ Theorem
to compute the conditional probability the Bulls won given there is a big crowd
overflowing into the parking lot. Do this.
Exercise 4.6 Assuming the probabilities in Example 4.12, compute the conditional probability of Joe having lung cancer given that he has a positive chest
X-ray.
Exercise 4.7 Suppose we have the Bayesian network in Figure 4.11. If we
instantiate X2 to 2 and X5 to 1, then A = {X2 , X5 } and a = {2, 1}. Suppose an
application of Algorithm 4.2 with m = 5 yields the following data:
Item
I1
I2
I3
I4
I5
I6
I7
I8
X1
1
1
2
1
2
2
1
2
X2
2
1
2
2
2
1
2
2
X3
1
X4
2
X5
1
1
2
1
2
1
2
1
1
2
1
2
2
1
1
1
Show the resultant estimates of the conditional probabilities of the remaining
variables.
Exercise 4.8 In Example 4.19, it was left as an exercise to compute the conditional probabilities of the remaining variables besides X1 . Do so.
Exercise 4.9 Suppose we have the Bayesian network in Figure 4.11. If we
instantiate X2 to 2 and X5 to 1, then A = {X2 , X5 } and a = {2, 1}. Suppose an
application of Algorithm 4.3 with m = 5 yields the following data:
Item
I1
I2
I3
I4
I5
X1
2
1
2
1
2
X2
2
2
2
2
2
X3
1
2
2
1
1
X4
2
1
2
1
2
X5
1
1
1
1
1
Compute the score of each item and the estimates of the conditional probabilities.
Section 4.3
Exercise 4.10 Sort the diseases in Example 4.22 in non-decreasing order according to their prior probabilities, and then apply Algorithm 4.4 to find a most
probable explanation. How many nodes were generated? Is it more or less than
the number of nodes generated when we did not sort them in Example 4.22?
238
CHAPTER 4. MORE INFERENCE ALGORITHMS
Chapter 5
Influence Diagrams
Consider again the Bayesian network in Figure 4.13. If a patient was, for example, a smoker and had a positive chest X-ray, a physician might consult that
network to determine how probable it was that the patient had lung cancer or
had bronchitis, or to determine the most probable explanation. The physician
would then use this and other information to arrive a decision as to how to
treat the patient. In general, the information obtained by doing inference in a
Bayesian network can be used to arrive at a decision even though the Bayesian
network itself does not recommend a decision. In this chapter, we extend the
structure of a Bayesian network so that the network actually does recommend
a decision. Such a network is called an influence diagram. Before discussing influence diagrams in Section 5.2, we present decision trees in Section 5.1, which
are mathematically equivalent to influence diagrams, and which are often simpler when the problem instance is small. After all this, Section 5.3 introduces
dynamic Bayesian networks and influence diagrams.
5.1
Decision Trees
After presenting some simple examples of decision trees, we discuss several issues
regarding their use.
5.1.1
Simple Examples
We start with the following example:
Example 5.1 Suppose your favorite stock NASDIP is down-graded by a reputable analyst and it plummets from $40 to $10 per share. You feel this is a good
buy, but there is a lot of uncertainty involved. NASDIP’s quarterly earnings are
about to be released and you think they will be good, which should positively influence its market value. However, you also think there is a good chance the whole
market will crash, which will negatively influence NASDIP’s market value. In
an attempt to quantify your uncertainty, you decide there is a .25 probability
239
240
CHAPTER 5. INFLUENCE DIAGRAMS
$5
$500
.25
Buy NASDIP
NASDIP
$10
.25
$1000
$20
.5
D
$2000
Leave $1000 in bank
$1005
Figure 5.1: A decision tree representing the problem instance in Example 5.1.
the market will crash, in which you case feel NASDIP will go to $5 by the end
of the month. If the market does not crash, you feel by the end of the month
NASDIP will be either at $10 or at $20 depending on the earnings report. You
think it is twice as likely it will be at $20 as at $10. So you assign a .5 probability
to NASDIP being at $20 and a .25 probability to it being at $10 at month end.
Your decision now is whether to buy 100 shares of NASDIP for $1000 or to
leave the $1000 in the bank where it will earn .005 interest in the next month.
One way to make your decision is to determine the expected value of your
investment if you purchase NASDIP and compare that value to the amount of
money you would have if you put the money in the bank. Let X be a random
variable, whose value is the worth of your $1000 investment in one month if you
purchase NASDIP. If NASDIP goes to $5, your investment will be worth $500,
if it stays at $10, your investment will be worth $1000, and if it goes to $20, it
will be worth $2000. Therefore,
E(X) = .25($500) + .25($1000) + .5($2000)
= $1375.
If you leave the money in the bank, your investment with be worth
1.005($1000) = $1005.
If you are what is called an ‘expected value maximizer’, your decision would
be to buy NASDIP.
The problem instance in the previous example can be represented by a decision tree. That tree is shown in Figure 5.1. A decision tree contains two kinds
of nodes: chance (or uncertainty) nodes representing random variables; and
decision nodes representing decisions to be made. We depict these nodes as
follows:
5.1. DECISION TREES
241
$5
$500
$1375
Buy NASDIP
NASDIP
.25
$10
.25
$1375
$1000
$20
.5
D
$2000
Leave $1000 in bank
$1005
Figure 5.2: The solved decision tree given the decision tree in Figure 5.1.
- chance node
- decision node
A decision represents a set of mutually exclusive and exhaustive actions the
decision maker can take. Each action is called an alternative in the decision.
There is an edge emanating from a decision node for each alternative in the
decision. In Figure 5.1, we have the decision node D with the two alternatives
‘Buy NASDIP’ and ‘Leave $1000 in bank.’ There is one edge emanating from a
chance node for each possible outcome (value) of the random variable. We show
the probability of the outcome on the edge and the utility of the outcome to
the right of the edge. The utility of the outcome is the value of the outcome to
the decision maker. When an amount of money is small relative to one’s total
wealth, we can usually take the utility of an outcome to be the amount of money
realized given the outcome. Currently, we make that assumption. Handling the
case where we do not make that assumption is discussed in Section 5.1.2. So,
for example, if you buy 100 shares of NASDIP and NASDIP goes to $20, we
assume the utility of that outcome to you is $2000. In Figure 5.1, we have
the chance node NASDIP with three possible outcome utilities, namely $500,
$1000, and $2000. The expected utility EU of a chance node is defined
to be the expected value of the utilities associated with its outcomes. The
expected utility of a decision alternative is defined to be the expected utility
of the chance node encountered if that decision is made. If there is certainty
when the alternative is taken, this expected utility is the value of that certain
outcome. So
EU (Buy NASDIP) = EU (N ASDIP ) = .25($500) + .25($1000) + .5($2000)
= $1375
242
CHAPTER 5. INFLUENCE DIAGRAMS
EU (Leave $1000 in bank) = $1005.
Finally, the expected utility of a decision node is defined to be the maximum of
the expected utilities of all its alternatives. So
EU (D) = max($1375, $1005) = $1375.
The alternative chosen is the one with the largest expected utility. The process
of determining these expected utilities is called solving the decision tree.
After solving it, we show expected utilities above nodes and an arrow to the
alternative chosen. The solved decision tree, given the decision tree in Figure
5.1, is shown in Figure 5.2.
Another example follows.
Example 5.2 Suppose you are in the same situation as in Example 5.1 except,
instead of considering leaving your money in the bank, your other choice is buy
an option on NASDIP. The option costs $1000, and it allows you to buy 500
shares of NASDIP at $11 per share in one month. So if NASDIP is at $5 or
$0 per share in one month, you would not exercise your option and you would
lose $1000. However, if NASDIP is at $20 per share in one month, you would
exercise your option, and your $1000 investment would be worth
500($20 − $11) = $4500.
Figure 5.3 shows a decision representing this problem instance. From that tree,
we have
EU (Buy option) = EU (N ASDIP2 ) = .25($0) + .25($0) + .5($4500)
= $2250.
Recall that EU (Buy NASDIP) is only $1375. So our decision would be to buy
the option. It is left as an exercise to show the solved decision tree.
Notice that the decision tree in Figure 5.3 is symmetrical, whereas the one
in Figure 5.2 is not. The reason is that we encounter the same uncertain event
regardless of which decision is made. Only the utilities of the outcomes are
different.
5.1.2
Probabilities, Time, and Risk Attitudes
Before proceeding, we address some concerns you may have. First, you may
be wondering how an individual could arrive at the probabilities of .25, .5, and
.25 in Example 5.1. These probabilities are not relative frequencies; rather they
are subjective probabilities that represent an individual’s reasonable numeric
beliefs. The individual arrives at them by a careful analysis of the situation.
Methods for assessing subjective beliefs were discussed briefly in Section 1.1.1
and are discussed in more detail in [Neapolitan, 1996]. Even so, you may argue
that the individual surely must believe there are many possible future values
5.1. DECISION TREES
243
$5
$500
.25
Buy NASDIP
NASDIP 1
$10
.25
$1000
$20
.5
D
Buy option
NASDIP 2
$2000
$5
.25
$0
$10
.25
$0
$20
.5
$4500
Figure 5.3: The decision tree modeling the investment decision concerning NASDIP, when the other choice is to buy an option on NASDIP.
for a share of NASDIP. How can the individual claim the only possible values
are $5, $10 and $20? You are correct. Indeed, this author is the individual in
this example, which concerns a recent investment decision of his. Although I
believe there are many possible future values, I feel the values $5, and $10, and
$20 with probabilities .25, .5, and .25 are sufficient to represent my beliefs as far
as influencing my decision. That is, I feel that further refinement of my beliefs
would not affect my decision.
Secondly, you many wonder why we based the decision on the outcome in
one month. Why not two months, a year, etc.? When using decision analysis
in problems such as these, the decision maker must base the decision on the
outcome at some point in the future. It is up to the decision maker’s preferences
to determine that point. This was my decision, and I based my decision on my
status one month into the future.
Finally, you may wonder why we chose the alternative with the largest expected value. Surely, a person who is very risk-averse might prefer the sure
$1005 over the possibility of ending up with only $500. This is absolutely true,
and it is possible to incorporate one’s attitude towards risk into the decision.
We modeled the situation where the utility of the outcome is the same as the
amount of money realized given the outcome. As mentioned previously, many
people maximize expected value when the amount of money is small relative
to their total wealth. The idea is that in the long run one will end up better
off by so doing. On the other hand, in the current example, I would not invest $100,000 in NASDIP because that is too much money relative to my total
wealth.
244
CHAPTER 5. INFLUENCE DIAGRAMS
1
0.8
0.6
0.4
0.2
0
1000
2000
x
3000
4000
5000
Figure 5.4: The U500 (x) = 1 − e−x/500 function.
1
0.8
0.6
0.4
0.2
0
1000
2000
x
3000
4000
5000
Figure 5.5: The U1000 (x) = 1 − e−x/1000 function.
One way to model an individual’s attitude towards risk is with a utility
function, which is a function that maps dollar amounts to utilities. An example
is the exponential utility function:
UR (x) = 1 − e−x/R .
In this function the parameter R, called the risk tolerance, determines how
risk-averse the function is. As R becomes smaller, the function becomes more
risk-averse. Figure 5.4 shows U500 (x), while Figure 5.5 shows U1000 (x). Notice
that the both functions are concave (opening downward) and the one in Figure
5.5 is closer to being a straight line. The more concave the more risk-averse the
function is, a straight line is risk neutral, and a convex (opening upward) function is risk-seeking. The following example illustrates the use of this function.
5.1. DECISION TREES
245
Example 5.3 Suppose we are making the decision in Example 5.1, and we use
the exponential utility function with R = 500. Then
EU (Buy NASDIP) = EU (N ASDIP )
= .25U500 ($500) + .25U500 ($1000) + .5U500 ($2000)
´
³
´
³
= .25 1 − e−500/500 + .25 1 − e−1000/500
´
³
+.5 1 − e−2000/500
= ..86504.
EU (Leave $1000 in bank) = U500 ($1005) = 1 − e−1005/500 = .86601.
If we modeled some individual’s attitude towards risk with this utility function,
that individual would choose to leave the money in the bank. It is left as an
exercise to show that using R = 1000 leads to the decision to buy NASDIP.
One way to determine your personal value of R is to consider a lottery in
which you will win $x with probability .5 and lose −$x/2 with probability .5.
Your value of R is the largest value of x for which you would choose the lottery
over doing nothing.
Modeling risk attitudes are discussed much more in [Clemen, 1996]. Henceforth, we simply assume the utility of an outcome is the same as the amount of
money realized given the outcome.
5.1.3
Solving Decision Trees
Next we show the general method for solving decision trees. There is a time
ordering from left to right in a decision tree. That is, any node to the right of
another node, occurs after that node in time. The tree is solved as follows:
Starting at the right,
proceed to the left
passing expected utilities to chance nodes;
passing maximums to decision nodes;
until the root is reached.
5.1.4
More Examples
We now present more complex examples of modeling with decision trees.
Example 5.4 Suppose Xia is a high roller and she is considering buying 10, 000
shares of ICK for $10 a share. This number of shares is so high that if she
purchases them, it could affect market activity and bring up the price of ICK.
She also believes the overall value of the DOW industrial average will affect the
price of ICK. She feels that, in one month, the DOW will either be at 10, 000
or 11, 000, and ICK will either be at $5 or $20 per share. Her other choice is
246
CHAPTER 5. INFLUENCE DIAGRAMS
$5
$50,000
.2
11,000
.6
ICK 1
$20
Buy ICK
.8
$200,000
DOW 1
$5
$50,000
.5
10,000
.4
ICK 2
$20
.5
D
$5
$200,000
$0
.3
11,000
.6
ICK 3
$20
Buy
option
.7
$250,000
DOW2
$5
.6
10,000
.4
$0
ICK 4
$20
.4
$250,000
Figure 5.6: A decision tree representing Xia’s decision as to whether to buy ICK
or an option on ICK.
5.1. DECISION TREES
247
to buy an option on ICK for $100, 000. The option will allow her to buy 50, 000
shares of ICK for $15 a share in one month. To analyze this problem instance,
she constructs the following probabilities:
P (ICK = $5|Decision = Buy ICK, DOW = 11, 000) = .2
P (ICK = $5|Decision = Buy ICK, DOW = 10, 000) = .5
P (ICK = $5|Decision = Buy option, DOW = 11, 000) = .3
P (ICK = $5|Decision = Buy option, DOW = 10, 000) = .6.
Furthermore, she assigns
P (DOW = 11, 000) = .6.
This problem instance is represented by the decision tree in Figure 5.6. Next we
solve the tree:
EU (ICK1 ) = (.2)($50, 000) + (.8)($200, 000) = $170, 000
EU (ICK2 ) = (.5)($50, 000) + (.5)($200, 000) = $125, 000
EU (Buy ICK) = EU (DOW1 ) = (.6)($170, 000) + (.4)($125, 000) = $152, 000
EU (ICK3 ) = (.3)($0) + (.7)($250, 000) = $175, 000
EU (ICK4 ) = (.6)($0) + (.4)($250, 000) = $100, 000
EU (Buy option) = EU (DOW2 ) = (.6)($175, 000) + (.4)($100, 000) = $145, 000
EU (D) = max($152, 000, $145, 000) = $152, 000.
The solved decision tree is shown in Figure 5.7. The decision is buy ICK.
The previous example illustrates a problem with decision trees. That is, the
representation of a problem instance by a decision tree grows exponentially with
the size of the instance. Notice that the instance in Examples 5.4 only has one
more element in it than the instance in Example 5.2. That is, it includes that
uncertainty about the DOW. Yet its representation is twice as large. So it is
quite difficult to represent a large instance with a decision tree. We will see in
the next section that influence diagrams do not have this problem. Before that,
we show more examples.
Example 5.5 Sam has the opportunity to buy a 1996 Spiffycar automobile for
$10, 000, and he has a prospect who would be willing to pay $11, 000 for the auto
if it is in excellent mechanical shape. Sam determines that all mechanical parts
except for the transmission are in excellent shape. If the transmission is bad, it
will cost Sam $3000 to repair it, and he would have to repair it before the prospect
would buy it. So he would only end up with $8000 if he bought the vehicle and
its transmission was bad. He cannot determine the state of the transmission
himself. However, he has a friend who can run a test on the transmission.
The test is not absolutely accurate. Rather 30% of the time it judges a good
transmission to bad and 10% of the time it judges a bad transmission to be
good. To represent this relationship between the transmission and the test, we
define the random variables which follow.
248
CHAPTER 5. INFLUENCE DIAGRAMS
$5
$170,000
11,000
.6
ICK 1
$152,000
Buy ICK
$20
.8
$200,000
DOW1
$5
$125,000
10,000
.4
$20
.5
D
$5
$175,000
11,000
.6
$50,000
.5
ICK 2
$152,000
$200,000
$0
.3
ICK 3
$20
$145,000
Buy
option
$50,000
.2
.7
$250,000
DOW 2
$100,000
10,000
.4
$5
.6
$0
ICK 4
$20
.4
$250,000
Figure 5.7: The solved decision tree given the decision tree in Figure 5.6.
5.1. DECISION TREES
249
$9722
Buy
Spiffycar
Tran1
$10,000
positive
.42
good
.571429
bad
D1
.428571
Do not buy
$11,000
$8000
$10,000
Test
Buy $10,897
Spiffycar
Tran2
$10,897
negative
.58
good
.965517
bad
D2
.034483
Do not buy
$11,000
$8000
$10,000
Figure 5.8: The decision tree representing the problem instance in Example 5.5.
Variable
T est
T ran
Value
positive
negative
good
bad
When the Variable Takes This Value
Test judges the transmission is bad
Test judges the transmission is good
Transmission is good
Transmission is good
The previous discussion implies we have these conditional probabilities:
P (T est = positive|T ran = good) = .3
P (T est = positive|T ran = bad) = .9.
Furthermore, Sam knows that 20% of the 1996 Spiffycars have bad transmissions. That is,
P (T ran = good) = .8.
Sam is going to have his friend run the test for free, and then he will decide
whether to buy the car.
This problem instance is represented in the decision tree in Figure 5.8. Notice
first that, if he does not buy the vehicle, the outcome is simply $10, 000. This
is because the point in the future is so near that we can neglect interest as
negligible. Note further that the probabilities in that tree are not the ones stated
in the example. They must be computed from the stated probabilities. We do
that next. The probability on the upper edge emanating from the T est node is
250
CHAPTER 5. INFLUENCE DIAGRAMS
$9722
Buy
Spiffycar
Tran1
$10,000
positive
.42
good
.571429
bad
D1
.429571
Do not buy
$11,000
$8000
$10,000
Test
Buy $10,897
Spiffycar
Tran2
$10,897
negative
.58
good
.965517
bad
D2
.034483
Do not buy
$11,000
$8000
$10,000
Figure 5.9: The solved decision tree given the decision tree in Figure 5.8.
the prior probability the test is positive. It is computed it as follows (Note that
we use our abbreviated notation.):
P (positive) = P (positive|good)P (good) + P (positive|bad)P (bad)
= (.3)(.8) + (.9)(.2) = .42.
The probability on the upper edge emanating from the T ran1 node is the probability the transmission is good given the test is positive. We compute it using
Bayes’ Theorem as follows:
P (good|positive) =
=
P (positive|good)P (good)
P (positive)
(.3)(.8)
= .571429.
.42
It is left as an exercise to determine the remaining probabilities in the tree.
Next we solve the tree:
EU (T ran1 ) = (.571429)($11, 000) + (.428571)($8000) = $9714
EU (D1 ) = max($9714, $10, 000) = $10, 000
EU(T ran2 ) = (.965517)($11, 000) + (.034483)($8000) = $10, 897
5.1. DECISION TREES
251
EU (D2 ) = max($10, 897, $10, 000) = $10, 897.
We need not compute the expected value of the T est node because there are no
decisions to the left of it. The solved decision tree is shown in Figure 5.9. The
decision is to not buy the vehicle if the test is positive and to buy it if the test
is negative.
The previous example illustrates another problem with decision trees. That
is, the probabilities needed in a decision tree are not always the ones that are
readily available to us. So we must compute them using the law of total probability and Bayes’ theorem. We will see that influence diagrams do not have this
problem either.
More examples follow.
Example 5.6 Suppose Sam is in the same situation as in Example 5.5 except
that the test is not free. Rather it costs $200. So Sam must decide whether to
run the test, buy the car without running the test, or keep his $10, 000. The
decision tree representing this problem instance is shown in Figure 5.10. Notice
that the outcomes when the test is run are all $200 less than their respective
outcomes in Example 5.5. This is because it cost $200 to run the test. Note
further that, if the vehicle is purchased without running the test, the probability
of the transmission being good is simply its prior probability .8. This is because
no test was run. So our only information about the transmission is our prior
information. Next we solve the decision tree. It is left as an exercise to show
EU (D1 ) = $9800
EU (D2 ) = $10, 697.
Therefore,
EU (T est) = (.42)($9800) + (.58)($10, 697) = $10, 320.
Furthermore,
EU (T ran3 ) = (.8)($11, 000) + (.2)($8000) = $10, 400.
Finally,
EU (D3 ) = max($10, 320, $10, 400, $10, 000) = $10, 400.
So Sam’s decision is to buy the vehicle without running the test. It is left as an
exercise to show the solved decision tree.
The following two examples illustrate cases in which the outcomes are not
numeric.
Example 5.7 Suppose Leonardo has just bought a new suit, he is about to leave
for work, and it looks like it might rain. Leonardo has a long walk from the
252
CHAPTER 5. INFLUENCE DIAGRAMS
good
Buy
Spiffycar
positive
.42
Tran1
bad
.428571
D1
Do not buy
Run test
$10,800
.571429
$7800
$9800
Test
good
Buy
Spiffycar
negative
.58
$10,800
.965517
Tran2
bad
D2
$7800
.034483
Do not buy
$9800
good
.8
Buy Spiffycar
D3
$11,000
Tran3
bad
$8000
.2
Do not buy
$10,000
Figure 5.10: The decision tree representing the problem instance in Example
5.6.
5.1. DECISION TREES
253
rain
suit ruined
.4
Do not
take umbrella
R
no rain
.6
D
Take umbrella
suit not ruined
suit not ruined,
inconveniece
Figure 5.11: The decision tree representing the problem instance in Example
5.7.
train to his office. So he knows if it rains and he does not have his umbrella,
his suit will be ruined. His umbrella will definitely protect his suit from the rain.
However, he hates the inconvenience of lugging the umbrella around all day.
Given he feels there is a .4 probability it will rain, should he bring his umbrella?
A decision tree representing this problem instance is shown in Figure 5.11. We
cannot solve this tree yet because its outcomes are not numeric. We can give
them numeric utilities as follows. Clearly, the ordering of the outcomes from
worst to best is as follows:
1. suit ruined
2. suit not ruined, inconvenience
3. suit not ruined.
We assign a utility of 0 to the worst outcome and a utility of 1 to the best
outcome. So
U (suit ruined) = 0
U (suit not ruined) = 1.
Then we consider lotteries (chance nodes) Lp in which Leonardo gets outcome
‘suit not ruined’ with probability p and outcome ‘suit ruined’ with probability
1 − p. The utility of ‘suit not ruined, inconvenience’ is defined to be the expected
utility of the lottery Lp0 for which Leonardo would be indifferent between lottery
Lp0 and being assured of ‘suit not ruined, inconvenience’. We then have
U (suit not ruined, inconvenience) ≡ EU (Lp0 )
= p0 U (suit not ruined)
+(1 − p0 )U (suit ruined)
= p0 (1) + (1 − p0 )0 = p0 .
254
CHAPTER 5. INFLUENCE DIAGRAMS
rain
0
suit ruined
1
suit not ruined
.8
suit not ruined,
inconveniece
.4
Do not
take umbrella
R
no rain
.6
D
Take umbrella
Figure 5.12: The decision tree with numeric values representing the problem
instance in Example 5.7.
Let’s say Leonardo decides p0 = .8. Then
U (suit not ruined, inconvenience) = .8
The decision tree with these numeric values is shown in Figure 5.12. We solve
that decision tree next:
EU (R) = (.4)(0) + (.6)(1) = .6
EU (D) = max(.6, .8) = .8.
So the decision is to take the umbrella.
The method used to obtain numeric values in the previous example easily
extends to the case where there are more than 3 outcomes. For example, suppose
there was a fourth outcome ‘suit goes to cleaners’ in between ‘suit not ruined,
inconvenience’ and ‘suit not ruined’ in the preference ordering. We consider
lotteries Lq in which Leonardo gets outcome ‘suit not ruined’ with probability q
and outcome ‘suit not ruined, inconvenience’ with probability 1 − q. The utility
of ‘suit goes to cleaners’ is defined to be the expected utility of the lottery Lq0
for which Leonardo would be indifferent between lottery Lq0 and being assured
of ‘suit goes to cleaners’. We then have
U (suit goes to cleaners) ≡ EU (Lq0 )
= q 0 U (suit not ruined)
+(1 − q 0 )U (suit not ruined, inconvenience)
= q 0 (1) + (1 − q 0 )(.8) = .8 + .2q 0 .
Let’s say Leonardo decides q 0 = .6. Then
U (suit goes to cleaners) = .8 + (.2)(.6) = .92.
5.1. DECISION TREES
255
no death
.999997
3 sore
throat days
Treat
A
anaphylaxis
death
.000003
D
Do not treat
dead
4 sore
throat days
Figure 5.13: A decision tree modeling Amit’s decision concerning being treated
for streptococcal infection.
Next we give an example from the medical domain.
Example 5.8 1 Amit, a 15 year old high school student, has been definitively
diagnosed with streptococcal infection, and he is considering having a treatment
which is known to reduce the number of days with a sore throat from 4 to 3. He
learns however that the treatment has a .000003 probability of causing death due
to anaphylaxis. Should he have the treatment?
You may argue that, if he may die from the treatment, he certainly should not
have it. However, the probability of dying is extremely small, and we daily accept
small risks of dying in order to obtain something of value to us. For example,
many people take a small risk of dying in a car accident in order to arrive at
work. We see then that we cannot discount the treatment based solely on that
risk. So what should Amit do? Next we apply decision analysis to recommend a
decision to him. Figure 5.13 shows a decision tree representing Amit’s decision.
To solve this problem instance we need to quantify the outcomes in that tree.
We can do this using quality adjusted life expectancies (QALE). We ask
Amit to determine what one year of life with a sore throat is worth relative to
one year of life without one. We will call such years ‘well years’. Let’s say he
says it is worth .9 well years. That is, for Amit
1 year with sore throat
is equivalent to
.9 well years.
We then assume a constant proportional trade-off. That is, we assume
the time trade-off associated with having a sore throat is independent of the
time spent with one (The validity of this assumption and alternative models are
1 This example is based on an example in [Nease and Owens, 1997]. Although the information is not fictitious, some of it is controversial.
256
CHAPTER 5. INFLUENCE DIAGRAMS
discussed in [Nease and Owens, 1997].). Given this assumption, for Amit
t years with sore throat
is equivalent to
.9t well years.
The value .9 is called the time-trade-off quality adjustment for a sore throat.
Another way to look at it is that Amit would give up .1 years of life to avoid
having a sore throat for .9 years of life. Now, if we let t be the amount of time
Amit will have a sore throat due to this infection, and l be Amit’s remaining life
expectancy, we define his quality QALE as follows:
QALE(l, t) = (l − t) + .9t.
From life expectancy charts, we determine Amit’s remaining life expectancy is
60 years. Converting days to years, we have the following:
3 days = .008219 years
4 days = .010959 years.
Therefore, Amit’s QALE’s are as follows:
QALE(60 yrs, 3 sore throat days) = 60 − .008219 + .9(.008219)
= 59.999178
QALE(60 yrs, 4 sore throat days) = 60 − .010959 + .9(.010959)
= 59.998904.
Figure 5.14 shows the decision tree in Figure 5.13 with the actual outcomes
augmented with QALE’s. Next we solve that tree:
EU (T reat) = EU (A) = (.999993)(59.999178) + (.000003)(0)
= 59.998758
EU (Do not treat) = 59.998904
EU (D) = max(59.998758, 59.998904) = 59.998904.
So the decision is to not treat, but just barely.
Example 5.9 This example is an elaboration of the previous one. Actually
streptococcus infection can lead to rheumatic heart disease (RHD), which is less
probable if the patient is treated. Specifically, if we treat a patient with streptococcus infection, the probability of rheumatic heart disease is .000013, while if
we do not treat the patient, the probability is .000063. The rheumatic heart disease would be for life. So Amit needs to take all this into account. First he must
determine time trade-off quality adjustments both for having rheumatic heart
5.1. DECISION TREES
257
no death
.999997
Treat
3 sore
throat days
59.999178 yrs
A
D
anaphylaxis
death
dead
.000003
0 yrs
4 sore
throat days
Do not treat
59.998904 yrs
Figure 5.14: The decision tree in Figure 5.13 with the actual oucomes augmented
by QALE’s.
disease alone and for having it along with a sore throat. Suppose he determines
the following:
1 year with RHD
is equivalent to
1 year with sore throat and RHD
.15 well years.
is equivalent to
.1 well years.
We then have
¶
µ
¶
µ
3
3
(.15) +
(.1)
QALE(60 yrs, RHD, 3 sore throat days) =
60 −
365
365
= 8.999589
µ
4
365
= 8.999452.
QALE(60 yrs, RHD, 4 sore throat days) =
60 −
¶
(.15) +
µ
¶
4
(.1)
365
We have already computed QALE’s for 3 or 4 days with only a sore throat in
the previous example. Figure 5.15 shows the resultant decision tree. We solve
that decision tree next:
EU (RHD1 ) = (.000013)(8.999569) + (.999987)(59.999178)
= 59.998515
EU (Treat) = EU(A) = (.999997)(59.998515) + (.000003)(0)
= 59.998335
258
CHAPTER 5. INFLUENCE DIAGRAMS
yes
.000013
no death
.999997
Treat
3 sore throat
days, RHD
8.999589 yrs
RHD1
no
A
.999987
3 sore throat
days, no RHD
59.999178 yrs
anayphylaxis death
dead
.000003
0 yrs
D
Do not
treat
yes
.000063
4 sore throat
days, RHD
8.999452 yrs
RHD2
no
.999937
4 sore throat
days, no RHD
59.998904 yrs
Figure 5.15: A decision tree modeling Amit’s decision concerning being treated
for streptococcal infection when rheumatic heart disease is considered.
EU (Do not treat) = EU (RHD2 )
= (.000063)(8.999452) + (.999937)(59.998904)
= 59.995691
EU (D) = max(59.998335, 59.995691) = 59.998335.
So now the decision is to treat, but again barely.
You may argue that, in the previous two examples, the difference in the
expected utilities is negligible because the number of significant digits needed to
express it is far more than the number of significant digits in Amit’s assessments.
This argument is reasonable. However, the utilities of the decisions are so close
because the probabilities of both anaphylaxis death and rheumatic heart disease
are so small. In general, this situation is not always the case. It is left as an
exercise to rework the previous example with the probability of rheumatic heart
disease being .13 instead of .000063.
Another consideration in medical decision making is the financial cost of
the treatments. In this case, the value of an outcome is a function of both the
QALE and the financial cost associated with the outcome.
5.2. INFLUENCE DIAGRAMS
5.2
259
Influence Diagrams
In the previous section, we noted the following two difficulties with decision
trees. First, the representation of a problem instance by a decision tree grows
exponentially with the size of the instance. Second, the probabilities needed in
a decision tree are not always the ones that are readily available to us. Next
we present an alternative representation of decision problem instances, namely
influence diagrams, which do not have either of these difficulties. First we discuss
representing problem instances with influence diagrams; then we discuss solving
influence diagrams.
5.2.1
Representing with Influence Diagrams
An influence diagram contains three kinds of nodes: chance (or uncertainty) nodes representing random variables; decision nodes representing
decisions to be made; and one utility node, which is a random variable whose
possible values are the utilities of the outcomes. We depict these nodes as
follows:
- chance node
- decision node
- utility node
The edges in an influence diagram have the following meaning:
Value of the node is probabilistically
dependent on the value of the parent.
Value of the parent is known at the time the decision
is made; hence the edge represents sequence.
Value of the node is deterministically
dependent on the value of the parent.
The chance nodes in an influence diagram satisfy the Markov condition with
the probability distribution. That is, for each chance node X, {X} is conditionally independent of the set of all its nondescendents given the set of all its
parents. So an influence diagram is actually a Bayesian network augmented
with decision nodes and a utility node. There must be an ordering of the decision nodes in an influence diagram based on the order in which the decisions
260
CHAPTER 5. INFLUENCE DIAGRAMS
P(NASDIP = $5) = .25
P(NASDIP = $10) = .25
P(NASDIP = $20) = .5
NASDIP
U
D
d1 = Buy NASDIP
d2 = Leave $1000 in bank
U(d1,$5) = $500
U(d1, $10) = $1000
U(d1, $20) = $2000
U(d2,n) = $1005
Figure 5.16: An influence diagram modeling your decision as to whether to buy
NASDIP.
are made. The order is specified using the edges between the decision nodes.
For example, if we have the order
D1 , D2 , D3 ,
then there are edges from D1 to D2 and D3 , and an edge from D2 to D3 .
To illustrate influence diagrams, we next represent the problem instances, in
the examples in the section on decision trees, by influence diagrams.
Example 5.10 Recall Example 5.1 in which you felt there is a .25 probability
NASDIP will be at $5 at month’s end, a .5 probability it will be at $20, and
a .25 probability it will be at $10. Your decision is whether to buy 100 shares
of NASDIP for $1000 or to leave the $1000 in the bank where it will earn .005
interest. Figure 5.16 shows an influence diagram representing this problem instance. Notice a few things about that diagram. There is no edge from D to
N ASDIP because your decision as to whether to buy NASDIP has no affect
on its performance (We assume your 100 shares is not enough to affect market
activity.). There is no edge from N ASDIP to D because at the time you make
your decision you do not know NASDIP’s value in one month. There are edges
from both N ASDIP and D to U because your utility depends both on whether
NASDIP goes up and whether you buy it. Notice that if you do not buy it, the
utility is the same regardless of what happens to NASDIP. This is why we write
U (d2, n) = $1005. The variable n represents any possible value of N ASDIP .
Example 5.11 Recall Example 5.2, which concerned the same situation as Example 5.1, except that your choices were either to buy NASDIP or to buy an
5.2. INFLUENCE DIAGRAMS
261
P(NASDIP = $5) = .25
P(NASDIP = $10) = .25
P(NASDIP = $20) = .5
NASDIP
D
d1 = Buy NASDIP
d2 = Buy option
U
U(d1,$5) = $500
U(d1, $10) = $1000
U(d1, $20) = $2000
U(d2,$5) = $0
U(d2,$10) = $0
U(d2,$20) = $4500
Figure 5.17: An influence diagram modeling your decision as to whether to buy
NASDIP when the other choice is to buy an option.
option on NASDIP. Recall further that if NASDIP is at $5 or $0 per share in
one month, you would not exercise your option and you would lose your $1000;
and, if NASDIP is at $20 per share in one month, you would exercise your
option and your $1000 investment would be worth $4500. Figure 5.17 shows an
influence diagram representing this problem instance. Recall that when we represented this instance with a decision tree (Figure 5.3) that tree was symmetrical
because we encounter the same uncertain event regardless of which decision is
made. This symmetry manifests itself in the influence diagram in that the value
of U depends on the value of the chance node N ASDIP regardless of the value
of the decision node D.
Example 5.12 Recall Example 5.4 in which Xia is a considering either buying
10, 000 shares of ICK for $10 a share, or an option on ICK for $100, 000 which
would allow her to buy 50, 000 shares of ICK for $15 a share in one month.
Recall further that she believes that, in one month, the DOW will either be at
10, 000 or at 11, 000, and ICK will either be at $5 or at $20 per share. Finally
recall that she assigns the following probabilities:
P (ICK = $5|DOW = 11, 000, Decision = Buy ICK) = .2
P (ICK = $5|DOW = 11, 000, Decision = Buy option) = .3
P (ICK = $5|DOW = 10, 000, Decision = buy ICK) = .5
262
CHAPTER 5. INFLUENCE DIAGRAMS
P(Dow=11,000) = .6
P(Dow=10,000) = .4
P(ICK=$5|Dow=11,000,D=d1) = .2
P(ICK=$5|Dow=11,000,D=d2) = .3
P(ICK=$5|Dow=10,000,D=d1) = .5
P(ICK=$5|Dow=10,000,D=d2) = .6
Dow
ICK
D
U
d1 = Buy ICK
d2 = Buy option
U(d1,$5) = $50,000
U(d1,$20) = $200,000
U(d2,$5) = $0
U(d2,$20) = $250,000
Figure 5.18: An influence diagram modeling Xia’s decision concerning buying
ICK or an option on ICK.
P (ICK = $5|DOW = 10, 000, Decision = Buy option) = .6
P (DOW = $11, 000) = .6.
Figure 5.18 shows an influence diagram representing this problem instance. Notice that the value of ICK depends not only on the value of the DOW but also
on the decision D. This is because Xia’s purchase can affect market activity.
Note further that this instance has one more component than the instance in
Example 5.11, and we needed to add only one more node to represent it with an
influence diagram. So the representation grew linearly with the size of the instance. By contrast, recall that, when we represented the instances with decision
trees, the representation grew exponentially.
Example 5.13 Recall Example 5.5 in which Sam has the opportunity to buy
a 1996 Spiffycar automobile for $10, 000, and he has a prospect who would be
willing to pay $11, 000 for the auto if it is in excellent mechanical shape. Recall
further that if the transmission is bad, Sam will have to spend $3000 to repair it
before he could sell the vehicle. So he would only end up with $8000 if he bought
the vehicle and its transmission was bad. Finally, recall he has a friend who can
run a test on the transmission, and we have the following:
P (T est = positive|T ran = good) = .3
P (T est = positive|T ran = bad) = .9
5.2. INFLUENCE DIAGRAMS
P(Test=positive| Tran=good) = .3
P(Test=positive| Tran=bad) = .9
263
P(Tran=good) = .8
Test
Tran
D
U
d1 = Buy Spiffycar
d2 = Do not buy
U(d1,good) = $11,000
U(d1,bad) = $8000
U(d2,t) = $10,000
Figure 5.19: An influence diagram modeling Sam’s decision concerning buying
the Spiffycar.
P (T ran = good) = .8.
Figure 5.19 shows an influence diagram representing this problem instance. Notice that there is an arrow from T ran to T est because the value of the test is
probabilistically dependent on the state of the transmission, and there is an arrow from T est to D because the outcome of the test will be known at the time
the decision is made. That is, D follows T est in sequence. Note further that the
probabilities in the influence diagram are the ones we know. We did not need to
use the law of total probability and Bayes’ Theorem to compute them, as we did
when we represented the instance with a decision tree.
Example 5.14 Recall Example 5.6 in which Sam is in the same situation as in
Example 5.5 except that the test is not free. Rather it costs $200. So Sam must
decide whether to run the test, buy the car without running the test, or keep
his $10, 000. Figure 5.20 shows an influence diagram representing this problem
instance. Notice that there is an edge from R to D because decision R is made
before decision D. Note further than again the representation of the instance
grew linearly with the size of the instance.
You may wonder why there is no edge from R to T est since the value of
T est is dependent on the decision R in the sense that the test will not be run
if Sam’s choice is r2 or r3. If a decision only affects whether the ‘experiment’
at a chance node takes place, and does not affect the outcome of the experiment
if it does take place, there is no need to draw an edge from the decision node
to the chance node. The reason is as follows: To each influence diagram there
corresponds a decision tree, which represents the same problem instance as the
influence diagram. By not including an edge from R to T est, we get a decision
tree that is symmetrical concerning the T est node rather than the one in Figure
264
CHAPTER 5. INFLUENCE DIAGRAMS
P(Test=positive| Tran=good) = .3
P(Test=positive| Tran=bad) = .9
R
r1 = Run test
r2 = Buy Spiffycar
r3 = Do not buy
P(Tran=good) = .8
P(Tran=bad) = .2
Test
Tran
D
U
d1 = Buy Spiffycar
d2 = Do not buy
U(r1,d1,good) = $10,800
U(r1,d1,bad) = $7800
U(r1,d2,t) = $9800
U(r2,d,good) = $11,000
U(r2,d, bad) = $8000
U(r3,d,t) = $10,000
Figure 5.20: An influence diagram modeling Sam’s decision concerning buying
the Spiffycar when he must pay for the test.
5.10. For the choices that do not run the test, the utilities of the outcomes will
be the same for both values of the T est node. So the solution to this decision
tree will be the same as the solution to the one in Figure 5.10. Contrast this
with the situation in Example 5.12, in which Xia’s decision does affect the value
of ICK. So we must have an arrow from the decision node D to the chance node
ICK.
Next we show a more complex instance, which we did not represent with a
decision tree.
Example 5.15 Suppose Sam is in the same situation as in Example 5.14, but
with the following modifications. First, Sam knows that 20% of the Spiffycars
were manufactured in a plant that produced lemons and 80% of them were manufactured in a plant that produced peaches. Furthermore, he knows 40% of the
lemons have good transmissions and 90% of the peaches have good transmissions. Also, 30% of the lemons have fine alternators and 80% of the peaches
have fine alternators. If the alternator is faulty (not fine), it will cost Sam $300
to repair it before he can sell he vehicle. Figure 5.21 shows an influence diagram
representing this problem instance. Notice that the set of chance nodes in the
influence diagram constitute a Bayesian network. For example, T ran and Alt
are not independent, but they are conditionally independent given Car.
We close with a large problem instance in the medical domain.
5.2. INFLUENCE DIAGRAMS
P(Test=positive|Tran=good) = .3
P(Test=positive|Tran=bad) = .9
R
r1 = Run test
r2 = Buy Spiffycar
r3 = Do not buy
265
P(Tran=good|Car=lemon) = .4
P(Tran=good|Car=peach) = .9
P(Car=lemon) = .2
P(Car=peach) = .8
Test
Tran
Car
D
U
Alt
d1 = Buy Spiffycar
d2 = Do not buy
U(r1,d1,good,fine) = $10,800
U(r1,d1,good,faulty) = $10,500
U(r1,d1,bad, fine) = $7800
U(r1,d1,bad,faulty) = $7500
U(r1,d2,t,a) = $9800
U(r2,d,good,fine) = $11,000
U(r2,d,good,faulty) = $10,700
U(r2,d,bad,fine) = $8000
U(r2,d, bad,faulty) = $7700
U(r3,d,t,a) = $10,000
P(Alt=fine|Car=lemon) = .3
P(Alt=fine|Car=peach) = .8
Figure 5.21: An influence diagram modeling Sam’s decision concerning buying
the Spiffycar when the alternator may be faulty.
Example 5.16 This example is taken from [Nease and Owens, 1997]. Suppose
a patient has a non-small-cell carcinoma of the lung. The primary tumor is 1
cm. in diameter, a chest X-ray indicates the tumor does not abut the chest wall
or mediastinum, and additional workup shows no evidence of distant metastases. The preferred treatment in this situation is thoracotomy. The alternative
treatment is radiation. Of fundamental importance in the decision to perform
thoracotomy is the likelihood of mediastinal metastases. If mediastinal metastases are present, thoracotomy would be contraindicated because it subjects the
patient to a risk of death with no health benefit. If mediastinal metastases are absent, thoracotomy offers a substantial survival advantage as long as the primary
tumor has not metastasized to distant organs.
We have two tests available for assessing the involvement of the mediastinum.
They are computed tomography (CT scan) and mediastinoscopy. This problem
instance involves three decisions. First, should the patient undergo a CT scan?
Second, given this decision and any CT results, should the patient undergo mediastinoscopy? Third, given these decisions and any test results, should the patient
undergo thoracotomy.
The CT scan can detect mediastinal metastases. The test is not absolutely
accurate. Rather, if we let M edM et be a variable whose values are present and
absent depending on whether or not mediastinal metastases are present, and
CT est be a variable whose values are cpos and cneg depending on whether or
266
CHAPTER 5. INFLUENCE DIAGRAMS
not the CT scan is positive, we have
P (CT est = cpos|M edMet = present) = .82
P (CT est = cpos|M edM et = absent) = .19.
The mediastinoscopy is an invasive test of mediastinal lymph nodes for determining whether the tumor has spread to those nodes. If we let Mtest be a
variable whose values are mpos and mneg depending on whether or not the
mediastinoscopy is positive, we have
P (M T est = mpos|M edM et = present) = .82
P (M T est = mpos|M edM et = absent) = .005.
The mediastinoscopy can cause death. If we let M be the decision concerning
whether to have mediastinoscopy, m1 be the choice to have it, and m2 be the
choice not to have it, and M edDeath be a variables whose values are mdie and
mlive depending on whether the patient dies from the mediastinoscopy, we have
P (M edDeath = mdie|M = m1) = .005
P (M edDeath = mdie|M = m2) = 0.
The thoracotomy has a greater chance of causing death than the alternative treatment radiation. If we let T be the decision concerning which treatment to have,
t1 be the choice to undergo thoracotomy, and t2 be the choice to undergo radiation, and T hordeath be a variables whose values are tdie and tlive depending
on whether the patient dies from the treatment, we have
P (T horDeath = tdie|T = t1) = .037
P (T horDeath = tdie|T = t2) = .002.
Finally, we need the prior probability that mediastinal metastases are present.
We have
P (M edM et = present) = .46.
Figure 5.22 shows an influence diagram representing this problem instance.
Note that we considered quality adjustments to life expectancy (QALE) and financial costs to be insignificant in this example. The value node is only in terms
of life expectancy.
5.2.2
Solving Influence Diagrams
We illustrate how influence diagrams can be solved using examples.
5.2. INFLUENCE DIAGRAMS
P(present ) = .46
Med
Met
CTest
U(t1,present ,tlive,mlive) = 1.8 yrs
U(t1,absent ,tlive, mlive) = 4.45 yrs
U(t2,present,tlive,mlive) = 1.8 yrs
U(t2,absent,tlive,mlive ) = 2.64 yrs
U(t,m,tdie,d) = 0
U(t,m,d,mdie ) = 0
MTest
P(mpos|present ) = .82
P(mpos|absent ) = .005
P(cpos|present ) = .82
P(cpos|absent ) = .19
t1 = Do thoracotomy
t2 = Do radiation
267
T
U
Thor
Death
P(tdie| t1) = .037
P(tdie| t2) = .002
C
c1 = Do CT
c2 = Do not do
M
m1 = Do mediastinoscopy
m2 = Do not do
Med
Death
P(mdie| m1) = .005
P(mdie| m2) = 0
Figure 5.22: An influence diagram modeling the decision as to whether to be
treated with thoracotomy.
Example 5.17 Consider the influence diagram in Figure 5.16, which was developed in Example 5.10. To solve the influence diagram, we need to determine
which decision choice has the largest expected utility. The expected utility of a
decision choice is the expected value E of U given the choice is made. We have
EU (d1) =
=
=
=
E(U |d1)
P ($5|d1)U (d1, $5) + P ($10|d1)U (d1, $10) + P ($20|d1)U (d1, $20)
(.25)($500) + (.25)($1000) + (.5)($2000)
$1375
EU (d2) =
=
=
=
E(U |d2)
P ($5|d2)U (d2, $5) + P ($10|d2)U (d2, $10) + P ($20|d2)U (d2, $20)
(.25)($1005) + (.25)($1005) + (.5)($1005)
$1005.
268
CHAPTER 5. INFLUENCE DIAGRAMS
The utility of our decision is therefore
EU (D) = max(EU (d1), EU (d2))
= max($1375, $1005) = $1375,
and our decision choice is d1.
Notice in the previous example that the probabilities do not depend on the
decision choice. This is because there is no edge from D to N ASDIP . In
general, this is not always the case as the next example illustrates.
Example 5.18 Consider the influence diagram in Figure 5.18, which was developed in Example 5.12. We have
EU (d1) =
=
=
=
E(U |d1)
P ($5|d1)U (d1, $5) + P ($20|d1)U (d1, $20)
(.32)($50, 000) + (.68)($200, 000)
$152, 000
EU (d2) =
=
=
=
E(U |d2)
P ($5|d2)U (d2, $5) + P ($20|d2)U (d2, $20)
(.42)($0) + (.58)($250, 000)
$145, 000
EU (D) = max(EU (d1), EU (d2))
= max($152, 000, $145, 000) = $152, 000,
and our decision choice is d1. You may wonder where we obtained the values
of P ($5|d1) and P ($5|d2). Once we instantiate the decision node, the chance
nodes comprise a Bayesian network. We then call a Bayesian network inference
algorithm to compute the needed conditional probabilities. For example, that
algorithm would do the following computation:
P ($5|d1) = P ($5|11, 000, d1)P (11, 000) + P ($5|10, 000, d1)P (10, 000)
= (.2)(.6) + (.5)(.4) = .32.
Henceforth, we will not usually show the computations done by the Bayesian
network inference algorithm. We will only show the results.
Example 5.19 Consider the influence diagram in Figure 5.19, which was developed in Example 5.13. Since there is an arrow from T est to D, the value
of T est will be known when the decision is made. So we need to determine the
5.2. INFLUENCE DIAGRAMS
269
expected value of U given each value of T est. We have
EU (d1| positive) = E(U |d1, positive)
= P (good|d1, positive)U (d1, good)
+P (bad|d1, positive)U (d1, bad)
= (.571429)($11, 000) + (.428571)($8000)
= $9714
EU (d2| positive) = E(U|d2, positive)
= P (good|d2, positive)U (d2, good)
+P (bad|d2, positive)U (d2, bad)
= (.571429)($10, 000) + (.428571)($10, 000)
= $10, 000
EU (D| positive) = max(EU (d1| positive), EU (d2| positive))
= max($9714, $10, 000) = $10, 000,
and our decision choice is d2. As in the previous example, the needed conditional
probabilities are obtained from a Bayesian network inference algorithm.
It is left as an exercise to compute EU (D| negative).
Example 5.20 Consider the influence diagram in Figure 5.20, which was developed in Example 5.14. Now we have two decisions, R and D. Since there is
an edge from R to D, decision R is made first and the EU of this decision is
the one we need to compute. We have
EU(r1) = E(U |r1)
= P (d1, good|r1)U (r1, d1, good) + P (d1, bad|r1)U (r1, d1, bad)
+P (d2, good|r1)U (r1, d2, good) + P (d2, bad|r1)U (r1, d2, bad)
We need to compute the conditional probabilities in this expression. Since D and
T ran are not dependent on R (Decision R only determines the value of decision
D in the sense that decision D does not take place for some values of R.), we
no longer show r1 to the right of the conditioning bar. We use an inference
algorithm for Bayesian networks to do these computations. For illustration, we
show them:
P (d1, good) = P (d1|good)P (good)
= [P (d1| positive)P (positive |good)
+P (d1| negative)P (negative |good)]P (good)
= [(0)P (positive |good) + (1)P (negative |good)] P (good)
= P (negative |good)P (good)
= (.7)(.8) = .56.
270
CHAPTER 5. INFLUENCE DIAGRAMS
The second equality above is obtained because D and T ran are independent
conditional on T est. The values of P (d1| positive) and P (d1| negative) were
obtained by first computing expected utilities as in Example 5.19, and then setting
the conditional probability to 1 if the decision choice is the one that maximizes
expected utility and to 0 otherwise. It is left as an exercise to show the other
three probabilities are .02, .24, and .18 respectively. We therefore have
EU (r1) = E(U |r1)
= P (d1, good)U (r1, d1, good) + P (d1, bad)U(r1, d1, bad)
+P (d2, good)U (r1, d2, good) + P (d2, bad)U (r1, d2, bad)
= (.56)($10, 800) + (.02)($7800) + (.24)($9800) + (.18)($9800)
= $10320.
It is left as an exercise to show
EU (r2) = $10, 400
EU (r3) = $10, 000.
So
EU (R) = max(EU (r1), EU (r2), EU (r2))
= max($10320, $10, 400, $10, 000) = $10, 500,
and our decision choice is r2.
Next we show another method for solving the influence diagram which, although may be less elegant than the previous method, corresponds more to the
way decision trees are solved. In this method, with decision R fixed at each of
its choices we solve the resultant influence diagram for decision D, and the we
use these results to solve R.
First fixing R at r1, we solve the influence diagram for D. The steps are
the same as those in Example 5.19. That is, since there is an arrow from T est
to D, the value of T est will be know when the decision is made. So we need to
determine the expected value of U given each value of T est. We have
EU (d1|r1, positive) = E(U |r1, d1, positive)
= P (good| positive)U (r1, d1, good)
+P (bad| positive)U (r1, d1, bad)
= (.571429)(11000) + (.429571)(8000)
= $9522
EU (d2|r1, positive) = E(U |r1, d2, positive)
= P (good| positive)U (r1, d2, good)
+P (bad| positive)U (r1, d2, bad)
= (.571429)($9800) + (.429571)($9800)
= $9800
5.2. INFLUENCE DIAGRAMS
271
EU (D|r1, positive) = max(EU(d1|r1, positive), EU (d2|r1, positive))
= max($9522, $9800) = $9800
EU (d1|r1, negative) = E(U |r1, d1, negative)
= P (good| negative)U (r1, d1, good)
+P (bad| negative)U (r1, d1, bad)
= (.965517)($10, 800) + (.034483)($7800)
= $10, 697
EU (d2|r1, negative) = E(U |r1, d2, negative)
= P (good| negative)U (r1, d2, good)
+P (bad| negative)U (r1, d2, bad)
= (.965517)($9800) + (.034483)($9800)
= $9800
EU(D|r1, negative) = max(EU (d1|r1, negative), EU (d2|r1, negative))
= max($10, 697, $9800) = $10, 697.
As before, the conditional probabilities are obtained from a Bayesian network
inference algorithm. Once we have the expected utilities of D, we can compute
the expected utility of R as follows:
EU (r1) = EU (D|r1, positive)P (positive) + EU (D|r1, negative)P (negative)
= $9800(.42) + $10, 697(.58)
= $10, 320.
Note that this is the same value we obtained using the other method. We next
proceed to compute EU (r2) and EU (r3) in the same way. It is left as an exercise
to do so.
The second method illustrated in the previous example extends readily to an
algorithm for solving influence diagrams. For example, if we had three decisions
nodes D1 , D2 , and D3 in that order, we would first instantiate D1 to its first
decision choice. Then, with D2 instantiated to its first decision choice, we’d
solve the influence diagram for D3 . We’d then compute the expected utility of
D2 ’s first decision choice. After doing this for all of D2 ’s decision choices, we’d
solve the influence diagram for D2 . We’d then compute the expected utility
of D1 ’s first decision choice. This process would be repeated for each of D1 ’s
decision choices. It is left an exercise to write an algorithm that implements
this method.
The algorithm just illustrated solves the influence diagram by converting it
on the fly to a decision tree. Shachter [1988] describes a way to evaluate an
272
CHAPTER 5. INFLUENCE DIAGRAMS
influence diagram without transforming it to a decision tree. The method operates directly on the influence diagram by performing arc reversal/node reduction
operations. These operations successively transform the diagram, ending with a
diagram with only one utility node that holds the utility of the optimal decision.
Shenoy [1992] describes another approach for evaluating influence diagrams.
The influence diagram is converted to a valuation network, and the nodes are
removed from this network by fusing the valuations bearing on the node to be removed. Shenoy’s algorithm is slightly more efficient than Shachter’s algorithm
because it maintains valuations, while Shachter’s algorithm maintains conditional probabilities. Additional operations are required to keep the probability
distributions normalized. Shachter and Ndilikijlikeshav[1993] modified the arc
reversal/node reduction algorithm to avoid these extra operations. The result is
an algorithm which has the same efficiency as Shenoy’s algorithm. Jensen et al
[1994] develop an algorithm which transforms an influence diagram to a junction
tree (See Section 3.4.). The algorithm is a based on the work in [Shenoy, 1992]
and [Jensen et al, 1990].
5.3
Dynamic Networks
After introducing dynamic Bayesian networks, we discuss dynamic influence
diagrams.
5.3.1
Dynamic Bayesian Networks
First we develop the theory; then we give an example.
Formulation of the Theory
Bayesian networks do not model temporal relationships among variables. That
is, a Bayesian network only represents the probabilistic relationships among a set
of variables at some point in time. It does not represent how the value of some
variable may be related to its value and the values of other variables at previous
points in time. In many problems, however, the ability to model temporal
relationships is very important. For example, in medicine it is often important
to represent and reason about time in tasks such as diagnosis, prognosis, and
treatment options. Capturing the dynamic (temporal) aspects of the problem is
also important in artificial intelligence, economics, and biology. Next we discuss
dynamic Bayesian networks, which do model the temporal aspects of a problem.
First we need define a random vector. Given random variables X1 , . . . and
Xn , the column vector


X1


X =  ... 
Xn
is called a random vector. A random matrix is defined in the same manner.
We use X to denote both a random vector and the set of random variables which
5.3. DYNAMIC NETWORKS
273
comprise X. Similarly, we use x to denote both a vector value of X and the set
of values that comprise x. The meaning is clear from the context. Given this
convention and a random vector X with dimension n, P (x) denotes the joint
probability distribution P (x1 , . . . xn ). Random vectors are called independent
if the sets of variables that comprise them are independent. A similar definition
holds for conditional independence.
Now we can define a dynamic Bayesian network, which extends the Bayesian
network to model temporal processes. We assume changes occur between discrete time points, which are indexed by the non-negative integers, and we have
some finite number T of points in time. Let {X1 , . . . Xn } be the set of features
whose values change over time, Xi [t] be a random variable representing the
value of Xi at time t for 0 ≤ t ≤ T , and let


X1 [t]


..
X[t] = 
.
.
Xn [t]
For all t, each Xi [t] has the same space which depends on i and we call it the
space of Xi . A dynamic Bayesian network is a Bayesian network containing
the variables that comprise the T random vectors X[t], and which is determined
by the following specifications:
1. An initial Bayesian network consisting of a) an initial DAG G0 containing
the variables in X[0] ; and b) an initial probability distribution P0 of these
variables.
2. A transition Bayesian network which is a template consisting of a) a transition DAG G→ containing the variables in X[t] ∪ X[t + 1]; and b) a
transition probability distribution P→ which assigns a conditional probability to every value of X[t + 1] given every value X[t]. That is, for every
value x[t + 1] of X[t + 1] and value x[t] of X[t] we specify
P→ (X[t + 1] = x[t + 1]|X[t] = x[t]).
Since for all t each Xi has the same space, the vectors x[t+1] and x[t] each
represent values from the same set of spaces. The index in each indicates
the random variable which has the value. We showed the random variables
above; henceforth we do not.
3. The dynamic Bayesian network containing the variables that comprise the
T random vectors consists of a) the DAG composed of the DAG G0 and
for 0 ≤ t ≤ T − 1 the DAG G→ evaluated at t; and b) the following joint
probability distribution:
P (x[0], . . . x[T ]) = P0 (x[0])
TY
−1
t=0
P→ (x[t + 1]|x[t]).
(5.1)
274
CHAPTER 5. INFLUENCE DIAGRAMS
X1[0]
X1[t]
X1 [t+1]
X1[0]
X1[1]
X1[2]
X2[0]
X2[t]
X2 [t+1]
X2[0]
X2[1]
X2[2]
X3[0]
X3[t]
X3 [t+1]
X3[0]
X3[1]
X3[2]
(b)
(a)
Figure 5.23: Prior and transition Bayesian networks are in (a). The resultant
dynamic Bayesian network for T = 2 is in (b). Note that the probablity distributions are not shown.
Figure 5.23 shows an example. The transition probability distribution entailed by the network in that figure is
P→ (x[t + 1]|x[t]) =
n
Y
i=0
P→ (xi [t + 1]|pai [t + 1]),
where pai [t + 1] denotes the values of the parents of Xi [t + 1]. Note that there
are parents in both X[t] and X[t + 1].
Owing to Equality 5.1, for all t and for all x
P (x[t + 1]|x[0], . . . x[t]) = P (x[t + 1]|x[t]).
That is, all the information needed to predict a world state at time t is contained
in the description of the world at time t − 1. No information about earlier times
is needed. Owing to this feature, we say the process has the Markov property.
Furthermore, the process is stationary. That is, P (x[t + 1]|x[t]) is the same
for all t. In general, it is not necessary for a dynamic Bayesian network to have
either of these properties. However, they reduce the complexity of representing
and evaluating the networks, and they are reasonable assumptions in many
applications. The process need not stop at an particular time T . However, in
practice we reason only about some finite amount of time. Furthermore, we
need a terminal time value to properly specify a Bayesian network.
Probabilistic inference in a dynamic Bayesian network can be done using
the standard algorithms discussed in Chapter 3. However, since the size of a
5.3. DYNAMIC NETWORKS
275
X1[0]
X1[t]
X1[t+1]
X1[0]
X1 [1]
X1[2]
E1[0]
E1[t]
E1[t+1]
E1[0]
E1 [1]
E1[2]
X2[0]
X2[t]
X2[t+1]
X2[0]
X2[1]
X2[2]
E2[0]
E2[t]
E2[t+1]
E2[0]
E2 [1]
E2[2]
X3[0]
X3[t]
X3[t+1]
X3[0]
X3[1]
X3 [2]
(a)
(b)
Figure 5.24: Prior and transition Bayesian networks, in the case where the networks in different time slots are connected only through non-evidence variables,
are in (a). The resultant dynamic Bayesian network for T = 2 is in (b).
dynamic Bayesian network can become enormous when the process continues
for a long time, the algorithms can be quite inefficient. There is a special
subclass of dynamic Bayesian networks in which this computation can be done
more efficiently. This subclass includes Bayesian networks in which the networks
in different time steps are connected only through non-evidence variables. An
example of such a network is shown in Figure 5.24. The variables labeled with an
E are the evidence variables and are instantiated in each time step. We lightly
shade nodes representing them. An application which uses such a dynamic
Bayesian network is shown in the next subsection. Presently, we illustrate how
updating can be done effectively in such networks.
Let e[t] be the set of values of the evidence variables at time step t and f[t]
be the set of of values of the evidence variables up to and including time step t.
Suppose for each value x[t] of X[t] we know
P (x[t]|f[t]).
We want to now compute P (x[t + 1]|f[t + 1]). First we have
276
CHAPTER 5. INFLUENCE DIAGRAMS
Figure 5.25: Tessellation of corrider layout.
P (x[t + 1]|f[t]) =
X
P (x[t + 1]|x[t], f[t])P (x[t]|f[t]).
x[t]
=
X
P (x[t + 1]|x[t])P (x[t]|f[t]).
(5.2)
x[t]
Using Bayes’ Theorem, we then have
P (x[t + 1]|f[t + 1]) = P (x[t + 1]|f[t], e[t + 1])
= αP (e[t + 1]|x[t + 1], f[t])P (x[t + 1]|f[t])
= αP (e[t + 1]|x[t + 1])P (x[t + 1]|f[t]),
(5.3)
where α is a normalizing constant. The value of P (e[t + 1]|x[t + 1]) can be
computed using an inference algorithm for Bayesian networks. We start the
process by computing P (x[0]|f[0]) = P (x[0]|e[0]). Then at each time step t + 1
we compute P (x[t + 1]|f[t + 1]) using Equalities 5.2 and 5.3 in sequence. Note
that to update the probability for the current time step we only need values
computed at the previous time step and the evidence at the current time step.
We can throw out all previous time steps, which means we need only keep
enough network structure to represent two time steps.
A simple way to view the process is as follows: We define
P 0 (x[t + 1]) ≡ P (x[t + 1]|f[t]),
which is the probability distribution of X[t + 1] given the evidence in the first t
time steps. We determine this distribution at the beginning of time step t + 1
using Equality 5.2, and then we discard all previous information. Next we obtain
the evidence in the time step t + 1 and update P 0 using Equality 5.3.
An Example: Mobile Target Localization
We show an application of dynamic Bayesian networks to mobile target localization, which was developed by Basye et al [1993]. The mobile target
5.3. DYNAMIC NETWORKS
277
580
574
5999
5999
217
5999
221
5999
Figure 5.26: Sonar readings upon entering a T-junction.
localization problem concerns tracking a target while maintaining knowledge
of one’s own location. Basye et al [1993] developed a world in which a target
and a robot reside. The robot is supplied with a map of the world, which is
divided into corridors and junctions. Figure 5.25 shows a portion of one such
world tessellated according to this scheme. Each rectangle in that figure is a
different region. The state space for the location of the target is the set of all
the regions shown in the figure, and the state space for the location of the robot
is the set of all these regions augmented with four quadrants to represent the
directions the robot can face. Let LR and LA be random variables whose values
are the locations of the robot and the target respectively.
Both the target and the robot are mobile, and the robot has sensors it uses
to maintain knowledge of its own location and to track the target’s location.
Specifically, the robot has a sonar ring consisting of 8 sonar transducers, configured in pairs pointing forward, backward, and to each side of the robot. Each
sonar gives a reading between 30 and 6000 millimeters, where 6000 means 6000
or more. Figure 5.26 shows one set of readings obtained from the sonars on
entering a T-junction. We want the sensors to tell us what kind of region we
are in. So we need a mapping from the raw sensor data to an abstract sensor
space consisting of the following: corridor, T-junction, L-junction, dead-end,
open space, and crossing. This mapping could be deterministic or probabilistic.
Basye et al [1993] discuss methods for developing it. Sonar data is notoriously
noisy and difficult to disambiguate. A sonar which happens to be pointed at an
angle of greater than 70 degrees to a wall, will likely not see that wall at all. So
we will assume the relationship is probabilistic. The robot also has a forward
pointing camera to identify the presence of its target. The camera can detect
the presence of a blob identified to be the target. If it does not detect a suitable
blob, this evidence is reported. If it does find a suitable blob, the size of the blob
is used to estimate its distance from the robot, which is reported in rather gross
units, i.e. within 1 meter, between 2 and 3 meters, etc. The detection of a blob
at a given distance is only probabilistically dependent on the actual presence of
278
CHAPTER 5. INFLUENCE DIAGRAMS
LR[0]
LR[t]
LR[t+1]
LR[0]
LR[1]
LR[2]
ER[0]
ER[t]
ER[t+1]
ER[0]
ER[1]
ER[2]
EA[0]
EA[t]
EA[t+1]
EA[0]
EA[1]
EA[2]
L A[0]
LA[t]
LA[t+1]
LA[0]
LA[1]
LA[2]
(a)
(b)
Figure 5.27: The prior and transition Bayesian networks for the mobile target
mobilization problem are (a). The resultant dynamic Bayesian network for
T = 2 is in (b).
the target at that distance. Let ER be a random variable whose value is the
sonar reading, which tells the robot something about its own location, and EA
be a random variable whose value is the camera reading, which tells the robot
something about the target’s location relative to the robot. It follows from the
previous discussion that ER is probabilistically dependent on LR , and EA is
probabilistically dependent on both LR and LA . At each time step, the robot
obtains readings from its sonar ring and camera. For example, it may obtain
the sonar readings in Figure 5.26, and its camera may inform it that the target
is visible at a certain distance.
The actions available to the robot and the target are as follows: travel down
the corridor the length of one region, turn left around the corner, turn around,
etc. In the dynamic Bayesian network model, these actions are simply performed
in some pre-programmed probabilistic way, which is not related to the sensor
data. So the location of the robot at time t + 1 is a probabilistic function of
its location at time t. When we model the problem with a dynamic influence
diagram in the Section 5.3.2, the robot will decide on its action based on the
sensor data. The target’s movement could be determined by a person or also
pre-programmed probabilistically.
In summary, the random variables in the problem are as follows:
5.3. DYNAMIC NETWORKS
Variable
LR
LA
ER
EA
279
What the Variable Represents
Location of the robot
Location of the target
Sensor reading regarding location of robot
Camera reading regarding location of target relative to robot
Figure 5.27 shows a dynamic Bayesian network which models this problem
(without showing any actual probability distributions). The prior probabilities
in the prior network represent information initially known about the location
of the robot and the target. The conditional probabilities in the transition
Bayesian network can be obtained from data. For example, P (eA |lR , lA ) can
obtained by repeatedly putting the robot and the target in positions lR and lA
respectively, and seeing how often reading eA is obtained.
Note that although the robot can sometimes view the target, the robot makes
no effort to track. That is, the robot moves probabilistically according to some
scheme. Our goal is for the robot to track the target. However, to do this
it must decide on where to move next based on the sensor data and camera
reading. As mentioned above, we need dynamic influences diagrams to produce
such a robot. They are discussed next.
5.3.2
Dynamic Influence Diagrams
Again we first develop the theory, and then we give an example.
Formulation of the Theory
To create a dynamic influence diagram from a dynamic Bayesian network we
need only add decision nodes and a value node. Figure 5.28 shows the high level
structure of such a network for T = 2. The chance node at each time step in that
figure represent the entire DAG at that time step, and so the edges represent
sets of edges. There is a edge from the decision node at time t to the chance
nodes at time t + 1 because the decision made at time t can affect the state
of the system at time t + 1. The problem is to determine the decision at each
time step which maximizes expected utility at some point in the future. Figure
5.28 represents the situation where we are determining the decision at time 0
which maximizes expected utility at time 2, The final utility, could in general,
be based on the earlier chance nodes and even the decision nodes. However, we
do not show such edges to simplify the diagram. Furthermore, the final expected
utility is often a weighted sum of expected utilities independently computed for
each time step up to the point in the future we are considering. Such a utility
function is called time-separable.
In general, dynamic influence diagrams can be solved using the algorithm
presented in Section 5.2.2. The next section contains an example.
280
CHAPTER 5. INFLUENCE DIAGRAMS
D[0]
D[1]
X[0]
X[1]
X[2]
U[2]
Figure 5.28: The high level structure of a dynamic influence diagram.
An Example: Mobile Target Localization Revisited
After we present the model, we show some results concerning a robot constructed
according to the model.
The Model Recall the robot discussed in Section 5.3.1. Our goal is for the
robot to track the target by deciding on its move at time t based on its evidence
at time t. So now we allow the robot to make a decision D[t] at time t as to
which action it will take, where the value of D[t] is a result of maximizing some
expected utility function based on the evidence in time step t. We assume there
is error in the robot’s movement. So the location of the robot at time t + 1 is
a probabilistic function of its location at the previous time step and the action
taken. The conditional probability distribution of LR is obtained from data as
discussed discussed at the end of Section5.3.1 That is, we repeatedly place the
robot in a location, perform an action, and then observe its new location.
The dynamic influence diagram, which represents the decision at time t and
in which the robot is looking three time steps into the future, is shown in Figure
5.29. Note that there are crosses through the evidence variable at time t to
indicate their values are already known. We need to maximize expected utility
using the probability distribution conditional on these values and the values of
all previous evidence variables. Recall at the end of Section 5.3.1 we called this
probability distribution P 0 and we discussed how it can be obtained. First we
need define a utility function. Suppose we decide to determine the decision at
time t by looking M time steps into the future. Let
dM = {d[t], d[t + 1], . . . d[t + M − 1]}
5.3. DYNAMIC NETWORKS
281
D[t]
D[t+1]
D[t+2]
LR[t]
LR[t+1]
LR[t+2]
LR[t+3]
ER[t]
ER[t+1]
ER[t+2]
ER[t+3]
EA[t]
EA[t+1]
EA[t+2]
EA[t+3]
LA[0]
LA[t+1]
LA[t+2]
LA[t+3]
Figure 5.29: The dynamic influence diagram modeling the robot’s decision as
to which action to take at time t.
be a set of values of the next M decisions including the current one and
fM = {eR [t + 1], eA [t + 1], eR [t + 2], eA [t + 2], . . . eR [t + M ], eA [t + M ]}
be a set of values of the evidence variables observed after the decisions are made.
For 1 ≤ k ≤ M let dk and fk respectively be the first k decisions and evidence
pairs in each of these sets. Define
X
Uk (fk , dk ) = − min
dist(u, v)P 0 (LA [t + k] = v)|fk , dk ),
(5.4)
u
v
where dist is the Euclidean distance, the sum is over all values v in the space of
LA , and the minimum is over all values u in the space of LA . Recall from the
beginning of Section 5.3.1 that the robot is supplied with a map of the world.
It uses this map to find every element in the space of LA .The idea is that if we
make these decisions and obtain these observations at time t + k, the sum in
Equality 5.4 is the expected value of the distance between the target and a given
location u. The smaller this expected value is the more likely it is the target is
close to u. The location ǔ which has the minimum expected value is then our
best guess at where the target is if we make these decisions and obtain these
observations. So the utility of the decisions and the observations is the expected
value for ǔ. The minus sign occurs because we maximize expected utility.
282
CHAPTER 5. INFLUENCE DIAGRAMS
We then have
EUk (dk ) =
X
fk
Uk (fk , dk )P 0 (fk |dk ).
(5.5)
This expected utility only concerns the situation k time steps into the future.
To take into account all time steps up to and including time t + M, we use a
utility function which is a weighted sum of the utilities at each time step. We
then have
M
X
γ k EUk (dk ),
(5.6)
EU (dM ) =
k=1
where γ k decreases with k to discount the impact of future consequences. Note
that implicitly γ k = 0 for k > M . Note further that we have a time-separable
utility function. We choose the decision sequence which maximizes this expected
utility in Equality 5.6, and we then make the first decision in this sequence at
time step t.
In summary, the process proceeds as follows: In time step t the robot updates
its probability distribution based on the evidence (sensor and camera readings)
obtained in that step. Then the expected utility of a sequence of decisions
(actions) is evaluated. This is repeated for other decision sequences, and the
one that maximizes expected utility is chosen. The first decision (action) in
that sequence is executed, the sensor and camera readings in time step t + 1 is
obtained, and the process repeats.
The computation of P 0 (fk |dk ) in Equality 5.5 for all values of f can be quite
expensive. Dean and Wellman [1991] discuss ways to reduce the complexity of
the decision evaluation.
Result: Emergent Behavior Basye et al [1993] developed a robot using
the model just described, and they observed some interesting, unanticipated
emergent behavior. By emergent behavior we mean behavior that is not
purposefully programmed into the robot, but that emerges as a consequence of
the model. For example, when the target moves towards a fork, the robot stays
close behind it, since this will enable it to determine which branch the target
takes. However, when the target moves towards a cul-de-sac, the robot keeps
fairly far away, whereas Basye et al [1993] expected it to remain close behind.
Analyzing the probability distributions and results of the value function, they
discovered that the model allows for the possibility that the target might slip
behind the robot, leaving the robot unable to determine the location of the target
without additional actions. If the robot stays some distance away, regardless of
what action the target takes, the observations made by the robot are sufficient
to determine the target’s location. Figure 5.30 illustrates the situation. In time
step t the robot is close to the target as the target is about to enter the cul-desac. If the robot stays close as illustrated by the top path, in time step t + 1 it
is just as likely that the target will slip behind the robot as it is that the target
will move up the cul-de—sac. If the target does slip behind the robot, it will no
longer be visible. However, if the robot backs off, as illustrated by the bottom
5.3. DYNAMIC NETWORKS
t
283
t+1
Figure 5.30: Staying close to the target may not be optimal.
path, the robot will be able determine the location of the target regardless of
what the target does. When considering its possible observations in time step
t + 1, the observation ‘target not visible’ would not give the robot a good idea
as to the target’s location. So the move to stay put is less valued than the move
to back off.
Larger-Scale Systems The method used to control our robot could be applied to a more complex system. Consider the following example taken from
[Russell and Norvig, 1995]. An autonomous vehicle uses a vision-based laneposition sensor to keep it in the center of its lane. The position sensor’s accuracy is directly affected by rain and an uneven road surface. Furthermore, both
rain and a bumpy road could cause the position sensor to fail. Sensor failure
of course affects the sensor’s accuracy. Two time steps in a dynamic influence
diagram, which models this situation, appears in Figure 5.31.
284
CHAPTER 5. INFLUENCE DIAGRAMS
D[t]
Laneposition[t]
Laneposition[t+1]
Positionsensor[t]
Positionsensor[t+1]
Sensoraccuracy[t]
Sensoraccuracy[t+1]
Weather[t]
Weather[t+1]
Sensorfailure[t]
Sensorfailure[t+1]
Terrain[t]
Terrain[t+1]
Figure 5.31: Two time steps in a dynamic influence diagram, which models the
decision faced by an autonomous vehicle.
EXERCISES
285
y1
.2
x1
100
Y
.7
y2
d1
200
.8
X
z1
300
.2
x2
Z
.3
z2
.5
z3
.3
D
d2
150
400
300
Figure 5.32: A decision tree.
Other Applications
Applications of dynamic Bayesian networks and influence diagrams include planning under uncertainty (e.g. our robot) [Dean and Wellman, 1991], analysis of
freeway traffic using computer vision [Huang et al, 1994], modeling the stepping patterns of the elderly to diagnose falls [Nicholson, 1996], and audio-visual
speech recognition [Nefian et al, 2002].
EXERCISES
Section 5.1
Exercise 5.1 Solve the decision tree in Figure 5.32.
Exercise 5.2 Solve the decision tree in Figure 5.33.
Exercise 5.3 Show the solved decision tree given the decision tree in Figure
5.3.
Exercise 5.4 Show that if we use R = 1000 in Example 5.3 the decision will
be to buy NASDIP.
286
CHAPTER 5. INFLUENCE DIAGRAMS
e1
x1
z1
.4
100
z2
.6
20
Z
E
.3
e2
d1
45
X
y1
220
w2
.2
10
W
.1
x2
w1
.8
Y
.7
y2
200
.9
v1
D
d2
190
.9
V
v2
.1
5
d3
150
Figure 5.33: A decision tree with two decisions.
Exercise 5.5 Compute the conditional probabilities in the tree in Figure 5.8
from the conditional probabilities given in Example 5.5.
Exercise 5.6 Show EU (D1 ) = $9820 and EU (D2 ) = $10, 697 for the decision
tree in Figure 5.10.
Exercise 5.7 Consider Example 5.7. Suppose Leonardo has the opportunity
to consult the weather forecast before deciding on whether to take his umbrella.
Suppose further that the weather forecast says it will rain on 90% of the days it
actually does rain, and 20% of the time it says it will rain on days it does not
rain. That is,
P (F orecast = rain|R = rain) = .9
EXERCISES
287
P (F orecast = rain|R = no rain) = .2.
As before, suppose Leonardo judges that
P (R = rain) = .4.
Show the decision tree representing this problem instance assuming the utilities
in Example 5.7. Solve that decision tree.
Exercise 5.8 Consider again Example 5.7. Assume if it rains, there is a .7
probability the suit will only need to go to the cleaners, and a .3 probability it
will rain. Assume again
P (R = rain) = .4.
Assess your own utilities for this situation, show the resultant decision tree, and
solve that tree.
Exercise 5.9 Consider Example 5.9. Assume your life expectancy from birth is
75 years. Assess your own QALE’s for the situation described in that example,
show the resultant decision tree, and solve that tree.
Exercise 5.10 Suppose Jennifer is a young, potential capitalist with $1000 to
invest. She has heard glorious tales of many who have made fortunes in the
stock market. So she decides to do one of three things with her $1000. She
could buy an option on Techjunk which would allow her to buy 1000 shares of
Techjunk for $22 a share in one month. She could use the $1000 to buy shares of
Techjunk. Finally, she could leave the $1000 in the bank earning .07% annually.
Currently, Techjunk is selling for $20 a share. Suppose further she feels there is
a .5 chance the NASDAQ will be at 1500 in two months and a .5 chance it will
be at 2000. If it is at 1500, she feels there is a .3 chance Techjunk will be at $23
a share and a .7 chance it will be at $15 a share. If the NASDAQ is at 2000,
she feels there is a .7 chance Techjunk will be at $26 a share and a .3 chance it
will be $20 a share. Show a decision tree that represents this decision and solve
that decision tree.
Let P (N ASDAQ = 3000) = p and P (N ASDAQ = 4000) = 1 − p. Determine the maximal value of p for which the decision would be to buy the option.
Is there any value of p for which the decision would be to buy the stock?
Exercise 5.11 This exercise is based on an example in [Clemen, 1996]. In
1984, Penzoil and Getty Oil agreed to a merger. However, before the deal was
closed, Texaco offered Getty a better price. So Gordon Getty backed out of the
Penzoil deal and sold to Texaco. Penzoil immediately sued, won the case, and
was awarded $11.1 billion. A court order reduced the judgment to $2 billion, but
interest and penalties drove the total back up to $10.3 billion. James Kinnear,
Texaco’s chief executive office, said he would fight the case all the way up to the
U.S. Supreme Court because, he argued, Penzoil had not followed Security and
Exchange Commission regulations when negotiating with Getty. In 1987, just
before Penzoil was to begin filing liens against Texaco, Texaco offered to give
288
CHAPTER 5. INFLUENCE DIAGRAMS
Penzoil $2 billion to settle the entire case. Hugh Liedke, chairman of Penzoil,
indicated that his advisors told him a settlement between $3 billion and $5 billion
would be fair.
What should Liedke do? Two obvious choices are 1) he could accept the
$2 billion; and 2) he could turn it down. Let’s say that he is also considering
counteroffering $5 billion. If he does, he judges that Texaco will either accept
the counteroffer with probability .17, refuse the counteroffer with probability .5,
or counter back in the amount of $3 billion with probability .33. If Texaco does
counter back, Liedke will then have the decision as to whether to refuse or accept
the counteroffer. Liedke assumes that if he simply turns down the $2 billion
with no counteroffer, or if Texaco refuses his counteroffer, or if he refuses their
return counteroffer, the matter will end up in court. If it does go to court,
he judges that there is .2 probability Penzoil will be awarded $10.3 billion, a
.5 probability they will be awarded $5 billion, and a .3 probability they will get
nothing.
Show a decision tree that represents this decision, and solve that tree.
What finally happened? Liedke simply refused the $2 billion. Just before
Penzoil began to file liens on Texaco’s assets, Texaco filed for protection from
creditors under Chapter 11 of the federal bankruptcy code. Penzoil then submitted a financial reorganization plan on Texaco’s behalf. Under the plan, Penzoil
would receive about $4.1 billion. Finally, the two companies agreed on $3 billion
as part of Texaco’s financial reorganization.
Section 5.2
Exercise 5.12 Solve the influence diagram in Figure 5.17, which was developed
in Example 5.11.
Exercise 5.13 Represent the problem instance in Exercise 5.7 with an influence
diagram. Solve that influence diagram.
Exercise 5.14 Represent the problem instance in Exercise 5.8 with an influence
diagram. Solve that influence diagram.
Exercise 5.15 Represent the problem instance in Exercise 5.10 with an influence diagram. Solve that influence diagram.
Exercise 5.16 Represent the problem instance in Exercise 5.11 with an influence diagram. Solve that influence diagram.
Section 5.3
EXERCISES
289
Exercise 5.17 Assign parameter values to the dynamic Bayesian network in
Figure 5.27, and compute the conditional probability of the locations of the robot
and the target at time 1 given certain evidence at times 0 and 1.
Exercise 5.18 Assign parameter values to the dynamic influence diagram in
Figure 5.29, and determine the decision at time 0 based on certain evidence at
time 0 and by looking 1 time into the future.
290
CHAPTER 5. INFLUENCE DIAGRAMS
Part III
Learning
291
Chapter 6
Parameter Learning:
Binary Variables
Initially the DAG in a Bayesian network was hand-constructed by a domain
expert. Then the conditional probabilities were assessed by the expert, learned
from data, or obtained using a combination of both techniques. For example
the DAGs in the Bayesian networks in Figures 1.11 and 3.1 were both handconstructed. However, the conditional probabilities in the Bayesian network in
Figure 3.1 were learned from data. Eliciting Bayesian networks from experts
can be a laborious and difficult procedure in the case of large networks. So researchers developed methods that could learn the DAG from data; furthermore,
they formalized methods for learning the conditional probabilities from data.
We present these methods in the chapters that follow. In a Bayesian network
the DAG is called the structure and the values in the conditional probability distributions are called the parameters. In this chapter and the next, we
address the problem of learning the parameter values from data. This chapter assumes each random variable has a space of size 2 (i.e. the variables are
binary.), while Chapter 7 concerns multinomial and continuous variables. Chapters 8-11 discuss learning the structure from data. Chapters 8 and 9 present a
Bayesian method for learning structure, while Chapter 10 presents an alternative method called constraint-based. Furthermore, in that chapter we show how
causal influences can be learned from data. Chapter 11 compares the methods
and presents several examples of learning both Bayesian networks and causal
influences.
We can only learn parameter values from data when the probabilities are
relative frequencies, which were discussed briefly in Section 1.1.1 and in more
detail in Section 4.2.1. Recall from this latter section that we often used the
term relative frequency rather than the term probability to refer to a propensity. That is the terminology used in this chapter. The word probability will
ordinarily refer to a subjective probability (degree of belief) as discussed in
Example 1.3. We will represent our belief concerning the value of a relative
293
294
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
frequency using a subjective probability distribution.
This chapter proceeds as follows: In Section 6.1 we discuss learning a single
parameter, which means we obtain an estimate of a single relative frequency.
Section 6.2 further discusses the Beta density function, which is introduced in
Section 6.1. Section 6.3 shows how to compute a probability interval for a
relative frequency, which enables us to express our confidence in its estimate.
Section 6.4 addresses the problem of learning all the parameters in a Bayesian
network. Learning parameters in the case of missing data items is covered in
Section 6.5. Finally, Section 6.6 shows a method for determining the variance
in the probability distribution of an inferred relative frequency in a Bayesian
network from the probability distributions of the parameters specified in the
network.
6.1
Learning a Single Parameter
After discussing subjective probability distributions of relative frequencies, we
develop a method for estimating a relative frequency from data.
6.1.1
Probability Distributions of Relative Frequencies
First we discuss developing a probability distribution of a relative frequency
when all number in [0, 1] are considered equally likely to be the relative frequency. Then we introduce a family of density functions which can be used to
represent an instance in which we do not feel all numbers in [0, 1] are equally
likely to be the relative frequency. Finally, we present the general method for
representing belief concerning a relative frequency using a subjective probability
distribution.
All Relative Frequencies Equally Probable
We present an urn example illustrating a probability distribution of a relative
frequency when all number in [0, 1] are equally likely to be the relative frequency,
First we discuss the case where the number of possible relative frequencies are
discrete; then we address the continuous case.
The Discrete Case Suppose we have 101 coins in a urn, each with a different
propensity for landing heads. The propensity for the first coin is .00, for the
second it is .01, for the third it is .02,... and for the last it is 1.00. This situation
is depicted in Figure 6.1. This means, if we tossed, for example, the second coin
many times, the relative frequency with which it landed heads would approach
.01. Suppose next that I pick a coin at random from the urn and I am about
to toss it. What probability should I assign to it landing heads? I think most
would agree that, if we knew the relative frequency with which the coin landed
heads, our probability of it landing heads would be that relative frequency. For
6.1. LEARNING A SINGLE PARAMETER
295
.98
.99
1.00
.03
.04
.05
.00
.01
.02
Figure 6.1: An urn containing 101 coins, each with a different propensity for
landing heads.
example, if we knew we were tossing the second coin, our probability1 would
be .01 because this is the relative frequency with which the second coin lands
heads. Let Side be a random variable whose values are the outcomes of the
toss, namely heads and tails, and F be a random variable whose range consists
of the 101 values of the relative frequencies.
P (Side = heads|f) = f.
Note that we used f to denote F = f , but we did not use shorthand notation for
Side = heads. This is consistent with the policy discussed in Section 1.1.4. If we
use the principle of indifference (See Section 1.1.1) to assign equal probabilities
to all relative frequencies (coins), we can represent our probability distribution
by the Bayesian network in Figure 6.2. Such a Bayesian network is called an
augmented Bayesian network because it includes a node representing our
belief about a relative frequency. Note that we shade that node. ‘Augmented
Bayesian network’ is formally defined in Section 6.4.2.We have then that
P (Side = heads) =
1.00
X
P (Side = heads|f )P (f )
f =.00
=
1.00
X
f =.00
=
µ
f
µ
1
101
1
100 × 101
¶
¶X
100
f =0
f=
µ
1
100 × 101
¶µ
100 × 101
2
¶
=
1
.
2
It is not surprising that the probability turns out to be .5 since the relative
1 Henceforth, I will simply refer to probabilities as ‘our probablity’, meaning only my belief.
It is simplest to state things that way. No one is compelled to agree.
296
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
P(f) = 1/101
.00
# f #1.00
F
Side
P(Side = heads| f) = f
Figure 6.2: A Bayesian network representing our belief concerning tossing a coin
picked at random from the urn in Figure 6.1
frequencies are distributed evenly on both sides of .5. What is this probability
value of .5? It is unlikely that it is the relative frequency with which the sampled
coin will land heads as that would be the case only if we picked the coin with
a propensity of .5. Rather it is our subjective probability (degree of belief) of
the coin landing heads on the first toss just as .6 was our subjective probability
of whether the Bulls won (discussed in Section 4.2.1). Similar to how we would
be indifferent between receiving a small prize if the Bulls won and receiving the
same small prize if a white ball was drawn from an urn containing 60% white
balls, we would be indifferent between receiving a small prize if the coin landed
heads and receiving the same small prize if a white ball was drawn from an urn
containing 50% white balls. Note that the value .5 is also the relative frequency
with which heads will occur if we repeatedly sample coins with replacement and
toss each sampled coin once.
The Continuous Case Suppose now that there is a continuum of coins in
the urn, for every real number f between 0 and 1 there is exactly one coin
with a propensity of f for landing heads, and we again pick a coin at random.
Then our probability distribution of the random variable, whose values are the
relative frequencies with which the coins land heads, is given by the uniform
density function. That is,
ρ(f ) = 1.
6.1. LEARNING A SINGLE PARAMETER
297
1
0.8
0.6
0.4
0.2
0
0.2
0.4 f 0.6
0.8
1
Figure 6.3: The uniform density function.
This density function is shown in Figure 6.3. In this case, our probability of
landing heads on the first toss is given by
Z 1
P (Side = heads|f )ρ(f)df
P (Side = heads) =
0
Z 1
1
=
f (1)df = .
2
0
Again, this result is not surprising.
Now consider some repeatable experiment such as the tossing of a thumbtack,
or sampling dogs and determining whether or not they eat the potato chips
which I offer them (I choose this example because I have no idea whether a
particular dog would eat potato chips.). If we feel all numbers in [0, 1] are
equally likely to be the value of the relative frequency, then we can model our
belief concerning the relative frequency using the uniform density function, just
as we did in the case of the coins in the urns.
All Relative Frequencies Not Equally Probable
In many, if not most, cases we do not feel all numbers in [0, 1] are equally likely
to be the value of a relative frequency. Even in the case of tossing a thumbtack,
I would not feel extreme values as probable as ones nearer the middle. More
notably, if I tossed a coin from my pocket, I would think it most probable that
the relative frequency is around .5. In this case, I would want a density function
similar to the one in Figure 6.4. If I sampled individuals in the United States and
298
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
8
6
4
2
0
0.2
0.4 f 0.6
0.8
1
Figure 6.4: The beta(f ; 50, 50) density function.
determined whether they brushed their teeth, I would think it most probably
the relative frequency is around .9. In this case, I would want a density function
like the one in Figure 6.5. Next we develop such density functions. Namely,
we discuss a family of density functions called the beta density functions, which
provide a natural way for quantifying prior beliefs about relative frequencies
and updating these beliefs in the light of evidence.
Before proceeding, we need review the gamma function, which is defined
as follows:
Z
∞
Γ(x) =
tx−1 e−t dt.
0
The integral on the right converges if and only if x > 0. If x is an integer ≥ 1,
it is possible to show
Γ(x) = (x − 1)!.
So the gamma function is a generalization of the factorial function. The following lemma concerns the gamma function:
Lemma 6.1 We have
Γ(x + 1)
= x.
Γ(x)
Proof. The proof is left as an exercise.
Now we can define the beta density function.
6.1. LEARNING A SINGLE PARAMETER
299
7
6
5
4
3
2
1
0
0.2
0.4 f 0.6
0.8
1
Figure 6.5: The beta(f ; 18, 2) density function.
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
0.2
0.4 f 0.6
0.8
1
Figure 6.6: The beta(f ; 3, 3) density function.
300
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
Definition 6.1 The beta density function with parameters a, b, N = a + b,
where a and b are real numbers > 0, is
ρ(f) =
Γ(N ) a−1
f
(1 − f )b−1
Γ(a)Γ(b)
0 ≤ f ≤ 1,
A random variable F , that has this density function, is said to have a beta
distribution.
We refer to the beta density function as beta(f ; a, b).
The uniform density function in Figure 6.3 is beta(f ; 1, 1). Figures 6.4, 6.5,
and 6.6 show other beta density functions. Notice that the larger the values of
a and b, the more the mass is concentrated around a/(a + b). For this and other
reasons, when a and b are integers, we often say the values of a and b are such
that the probability assessor’s experience is equivalent to having seen the first
outcome occur a times in a + b trials. The results in Section 6.1.2 give further
justification for this statement.
We will need the following two lemmas concerning the beta density function:
Lemma 6.2 If a and b are real numbers > 0 then
Z 1
Γ(a + 1)Γ(b + 1)
.
f a (1 − f )b df =
Γ(a + b + 2)
0
Proof. The proof is left as an exercise.
Lemma 6.3 If F has a beta distribution with parameters a, b, N = a + b, then
E(F ) =
a
.
N
Proof. We have
E(F ) =
Z
1
f ρ(f )df
0
=
=
=
=
=
Z
1
Γ(N ) a−1
f
(1 − f )b−1 df
Γ(a)Γ(b)
0
Z 1
Γ(N )
f a (1 − f )b−1 df
Γ(a)Γ(b) 0
Γ(N )
Γ(a + 1)Γ(b)
Γ(a)Γ(b) Γ(a + b − 1 + 2)
Γ(N ) Γ(a + 1)Γ(b)
Γ(a)Γ(b) Γ(N + 1)
a
.
N
f
The fourth equality above is due to Lemma 6.2 and the last is due to Lemma
6.1.
6.1. LEARNING A SINGLE PARAMETER
301
k(f)
F
X
P(X = 1|F = f) = f
Figure 6.7: The probability distribution of F represents our belief concerning
the relative frequency with which X equals 1.
Representing Belief Concerning a Relative Frequency
Next we formalize and generalize the notions introduced in Section 6.1.1. Suppose we have some two-outcome random process such as the tossing of a thumbtack, and we let X be a random variable whose values, 1 and 2, are the outcomes
of the experiment2 . We assume we can represent our belief concerning the relative frequency with which X equals 1 using a random variable F whose space
consists of numbers in the interval [0, 1]3 . The expected value E(F ) is defined
to be our estimate of the relative frequency. We further assume our beliefs
are such that
P (X = 1|f ) = f.
(6.1)
That is, if we knew for a fact that the relative frequency with which X equals
1 was f , our belief concerning the occurrence of 1 in the first execution of the
experiment would be f . This situation is represented by the Bayesian network
in Figure 6.7. Given this assumption, the theorem that follows shows that our
subjective probability for the first trial is the same as our estimate of the relative
frequency.
Theorem 6.1 Suppose X is a random variable with two values 1 and 2, F is
another random variable such that
P (X = 1|f ) = f.
2 Recall in previous chapters we ordinarily used x1 and x2 as the values of variable X when
we do not use names that have semantic connotations. For the sake of notational simplicity,
in this chapter we simply use 1 and 2 as the values of all variables.
3 It is somewhat standard to use theta (θ) for a random variable whose value is a relative
frequency. However, owing to the similarity of capital and small theta, we find it more lucid
to use F for relative frequency.
302
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
Then
P (X = 1) = E(F ),
where E stands for expected value.
Proof. We prove the case where F is continuous. Owing to the law of total
probability,
Z 1
P (X = 1|f )ρ(f )df
P (X = 1) =
0
Z 1
=
fρ(f )df
0
= E(F ).
Corollary 6.1 If the conditions in Theorem 6.1 hold, and F has a beta distribution with parameters a, b, N = a + b, then
P (X = 1) =
a
.
N
Proof. The proof follows immediately from Theorem 6.1 and Lemma 6.3.
Example 6.1 Suppose I am going to repeatedly toss a coin from my pocket.
Since I would feel it highly probable that the relative frequency is around .5,
I might feel my prior experience is equivalent to having seen 50 heads in 100
tosses. Therefore, I could represent my belief concerning the relative frequency
with the beta(f ; 50, 50) density function, which is shown in Figure 6.4. Due to
the previous corollary, for the first toss of the coin,
P (Side = heads) =
50
= .5.
50 + 50
Furthermore, .5 is our estimate of the relative frequency with which the coin will
land heads.
Example 6.2 Suppose I am going to repeatedly toss a thumbtack. Based on
its structure, I might feel it should land heads about half the time, but I would
not be nearly so confident as I would with the coin from my pocket. So I might
feel my prior experience is equivalent to having seen 3 heads (landing on its flat
side) in 6 tosses, which means I could represent my belief concerning the relative
frequency with the beta(f ; 3, 3) density function, which is shown in Figure 6.6.
Due to the previous corollary, for the first toss of the thumbtack,
P (Side = heads) =
3
= .5.
3+3
Furthermore, .9 is our estimate of the relative frequency of individuals who brush
their teeth.
6.1. LEARNING A SINGLE PARAMETER
303
Example 6.3 Suppose I am going to sample individuals in the United States
and determine whether they brush their teeth. In this case, I might feel my prior
experience is equivalent to having seen 18 individuals brush their teeth out of 20
sampled. Therefore, I could represent my belief concerning the relative frequency
with the beta(f ; 18, 2) density function, which is shown in Figure 6.5. Due to
the previous corollary, for the first individual sampled,
P (T eeth = brushed) =
18
= .9.
18 + 2
Furthermore, .9 is our estimate of the relative frequency of individuals who brush
their teeth.
6.1.2
Learning a Relative Frequency
Recall the coins in the urn in Figure 6.1. We determined that, if we picked
a coin at random and tossed it, our probability of landing heads on that toss
would be .5. Suppose now that we’ve tossed it 20 times and it landed heads 18
times. Would we still assign a probability of .5 to the next toss? We would not
because we would now feel it more probable that the coin is one of the coins
with propensity around .9 than we would that it is one of the coins with a small
propensity. Next we discuss how to quantify such a change in belief.
Suppose we perform M trials of a random process. Let X (h) be a random
variable whose value is the outcome of the hth trial, and let F be a random
variable whose probability distribution represents our belief concerning the relative frequency. We assume that if we knew the value of F for certain, then
we would feel the X (h) s are mutually independent, and our probability for each
trial would be that relative frequency. That is, if, for example, we were tossing
a coin whose propensity we knew to be .5, then our probability for the hth toss
would be .5 regardless of the outcomes of the first h − 1 tosses.
Why should we make the assumption in the previous paragraph? First, there
is evidence that separate trials are independent as far as the actual relative
frequencies are concerned. That is, as discussed in Section 4.2.1, in 1946 J.E.
Kerrich conducted many experiments indicating the relative frequency appears
to approach a limit; in 1971 G.R. Iversen et al ran many experiments with
dice indicating that what we call random processes do indeed generate random
sequences; and in 1928 R. von Mises proved that separate trials are independent
if we assume the relative frequency approaches a limit and a random sequence
is generated.
You may argue that, even though there is evidence that separate trials are
independent as far as the actual relative frequencies are concerned, we are not
compelled to say that, given we know the relative frequencies, they are independent as far as our beliefs are concerned. This may be true. However, Bruno de
Finetti showed that if we make certain minimal assumptions about our beliefs,
then we must believe they are independent given some random variable. He
assumes exchangeability, which means an individual assigns the same probability to all sequences of the same length containing the same number of each
304
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
k(f)
F
X(1)
X(2)
X(M)
P(X(1) = 1| f) = f
P(X(2) = 1| f) = f
P(X(M) = 1| f) = f
Figure 6.8: A Bayesian network representing the fact that the X (h) s are independent conditional on F .
outcome. For example, if we denote a heads by a 1 and a tails by 2, the individual assigns the same probability to these two sequences:
1212121222
and
1221122212.
Furthermore, this same probability is assigned to all other sequences of length
ten that have precisely four 1’s and six 0’s. De Finetti’s assumption of exchangeability is similar to von Mises’ assumption that the sequence of outcomes is a
random sequence (See Section 4.2.1.). However, exchangeability has to do with
an individual’s beliefs, whereas randomness has to do with an objective property of nature. Exchangeability could serve as a Bayesian definition of ‘random
process’. That is, a repeatable experiment is considered a random process
by an individual if the individual’s beliefs (probabilities) concerning sequences
of outcomes of the experiment satisfy exchangeability. Given the assumption
of exchangeability, in 1937 B. de Finetti proved there must be some random
variable that renders the individual’s beliefs concerning the trials independent.
These matters are discussed more in [Good, 1965].
Given all of the above, we will assume the X (h) s are mutually independent
conditional on the value of the relative frequency. This independence is represented by the Bayesian network in Figure 6.8.
Next we give formal definitions concerning the notion just developed.
Definition 6.2 Suppose
1. We have a set of random variables (or random vectors) D = {X (1) , X (2) , . . .
X (M ) } such that each X (h) has the same space;
2. There is a random variable F with density function ρ such that the X (h) s
are mutually independent conditional on F, and for all values f of F, all
X (h) have the same probability distribution conditional on f.
6.1. LEARNING A SINGLE PARAMETER
305
Then D is called a sample of size M with parameter F.
Given a sample, the density function ρ is called the prior density function of the parameters relative to the sample. It represents our prior belief
concerning the unknown parameters. Given a sample, the marginal distribution
of each X (h) is the same. This distribution is called the prior distribution
relative to the sample. It represents our prior belief concerning each trial.
In general, F is a set containing more than one element, at least one of which
is a random variable. Furthermore, the members of D may be random vectors.
Such a case is discussed in Section 6.4.3 (Rrandom vectors were defined Section
5.3.1.). However, in the case of binomial samples, which are discussed next, F
contains only one element and the members of D are random variables.
Definition 6.3 Suppose we have a sample of size M such that
1. each X (h) has space {1, 2};
2. F = {F }, F has space [0, 1], and for 1 ≤ h ≤ M
P (X (h) = 1|f ) = f.
(6.2)
Then D is called a binomial sample of size M with parameter F .
Example 6.4 Suppose we sample a coin from the urn in Figure 6.1 and toss it
twice. Let 1 be the outcome if a heads occurs, 2 be the outcome if a tails occurs,
and X (h) ’s value be the outcome of the hth toss. Clearly, the density function
for F is beta(f ; 1, 1). So D = {X (1) , X (2) } is a binomial sample of size 2 with
parameter F . Owing to Theorem 6.1 and Corollary 6.1, the prior distribution
relative to the sample is
P (X (h) = 1) = E(F ) =
1
1
= .
1+1
2
Example 6.5 Suppose we toss a thumbtack 10 times. Let 1 be the outcome if a
heads occurs, 2 be the outcome if a tails occurs, and X (h) ’s value be the outcome
of the hth toss. Furthermore, let the density function for F be beta(f ; 3, 3). So
D = {X (1) , X (2) , . . . X (10) } is a binomial sample of size 10 with parameter F .
Owing to Theorem 6.1 and Corollary 6.1, the prior distribution relative to the
sample is
1
3
= .
P (X (h) = 1) = E(F ) =
3+3
2
Before developing theory that enables us to update our belief about the next
trial from a binomial sample, we present two more lemmas concerning the beta
distribution.
Lemma 6.4 Suppose F has a beta distribution with parameters a, b, N = a + b,
s and t are two integers ≥ 0, and M = s + t. Then
¢
¡
E F s [1 − F ]t =
Γ(N ) Γ(a + s)Γ(b + t)
.
Γ(N + M )
Γ(a)Γ(b)
306
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
Proof. We have
Z
¡
¢
E F s [1 − F t ] =
1
f s (1 − f)t ρ(f)df
0
Z
=
1
f s (1 − f)t
0
Γ(N ) a−1
f
(1 − f )b−1 df
Γ(a)Γ(b)
Z 1
Γ(N )
f a+s−1 (1 − f)b+t−1 df
Γ(a)Γ(b) 0
Γ(N ) Γ(a + s)Γ(b + t)
Γ(a)Γ(b) Γ(a + s + b + t)
Γ(N ) Γ(a + s)Γ(b + t)
.
Γ(N + M )
Γ(a)Γ(b)
=
=
=
The fourth equality above is due to Lemma 6.2.
Lemma 6.5 Suppose F has a beta distribution with parameters a, b, N = a + b,
and s and t are two integers ≥ 0. Then
f s (1 − f )t ρ(f )
= beta(f ; a + s, b + t).
E(F s [1 − F ]t )
Proof. Let M = s + t. We have
f s (1 − f )t ρ(f )
E(F s [1 − F ]t )
=
Γ(N )
f s (1 − f )t Γ(a)Γ(b)
f a−1 (1 − f )b−1
Γ(N ) Γ(a+s)Γ(b+t)
Γ(N +M )
Γ(a)Γ(b)
Γ(N + M )
f a+s−1 (1 − f )b+t−1
Γ(a + s)Γ(b + t)
= beta(f ; a + s, b + t).
=
The first equality above is due Lemma 6.4.
We now obtain results enabling us to update our belief from a binomial
sample.
Theorem 6.2 Suppose
1. D is a binomial sample of size M with parameter F ;
2. we have a set of values
d = {x(1) , x(2) , . . . x(M) }
of the variables in D (The set d is called our data set (or simply data).);
3. s is the number of variables in d equal to 1;
4. t is the number of variables in d equal to 2.
6.1. LEARNING A SINGLE PARAMETER
307
Then
P (d) = E(F s [1 − F ]t ).
Proof. We prove the case where F is continuous. We have
Z 1
P (d) =
P (d|f )ρ(f )df
0
=
=
Z
M
1 Y
0 h=1
1
s
Z
0
P (x(h) |f )ρ(f )df
f (1 − f )t ρ(f)df
= E(F s [1 − F ]t ).
The second equality above is because the variables in D are conditionally independent conditional on F , and the third is obtained by repeatedly applying
Equality 6.2 in Definition 6.3.
Recall from Section 4.2.1 that we also use the word ‘sample’ for the set of
observations, which is our data set d, and that it is clear from the context which
we mean.
Corollary 6.2 If the conditions in Theorem 6.2 hold, and F has a beta distribution with parameters a, b, N = a + b, then
P (d) =
Γ(N ) Γ(a + s)Γ(b + t)
.
Γ(N + M )
Γ(a)Γ(b)
Proof. The proof follows immediately from Theorem 6.2 and Lemma 6.4.
Example 6.6 Suppose we have the binomial sample in Example 6.4, and
d = {1, 2}.
Then a = b = 1, N = 2, s = 1, t = 1, M = 2, and due to the preceding theorem,
P (d) =
1
Γ(2) Γ(1 + 1)Γ(1 + 1)
= .
Γ(2 + 2)
Γ(1)Γ(1)
6
Similarly, if d0 = {1, 1}, P (d0 ) = 1/3. You may wonder why the probability of
two heads is twice the probability of a heads followed by a tail. Note that
P (d) = P (X (1) = 1, X (2) = 2) = P (X (2) = 2|X (1) = 1)P (X (1) = 1)
while
P (d0 ) = P (X (1) = 1, X (2) = 1) = P (X (2) = 1|X (1) = 1)P (X (1) = 1).
Intuitively, P (X (2) = 1|X (1) = 1) is greater than P (X (2) = 2|X (1) = 1) because
once the first toss results in heads, it becomes more likely we have a coin with
propensity for landing heads. So the probability of heads on the second toss is
increased.
308
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
Note that in the previous example, P (d) is the relative frequency with which
we will obtain data d when we repeatedly sample a coin with replacement from
the urn and toss it twice.
Example 6.7 Suppose we have the binomial sample in Example 6.5, and
d = {1, 1, 2, 1, 1, 1, 1, 1, 2, 1}.
Then a = b = 3, N = 6, s = 8, t = 2, M = 10, and due to the preceding
corollary,
Γ(6) Γ(3 + 8)Γ(3 + 2)
= .001998.
P (d) =
Γ(6 + 10)
Γ(3)Γ(3)
Theorem 6.3 If the conditions in Theorem 6.2 hold, then
ρ(f|d) =
f s (1 − f )t ρ(f )
E(F s [1 − F ]t )
where ρ(f |d) denotes the density function of F conditional on D = d.
Proof. We have
ρ(f |d) =
=
P (d|f)ρ(f )
P (d)
s
f (1 − f )t ρ(f )
.
E(F s [1 − F ]t )
The first equality above is due to Bayes’ Theorem. The second equality is due
to the fact that the variables in D are independent conditional on F , Equality
6.2, and Theorem 6.2.
Corollary 6.3 Suppose the conditions in Theorem 6.2 hold, and F has a beta
distribution with parameters a, b, N = a + b. That is,
ρ(f ) = beta(f ; a, b).
Then
ρ(f |d) = beta(f ; a + s, b + t).
Proof. The proof follows immediately from Theorem 6.3 and Lemma 6.5.
Given a sample, the density function denoted ρ|d is called the updated
density function of the parameters relative to the sample. It represents
our posterior belief concerning the unknown parameters. The previous corollary
shows that when we update a beta density function relative to a binomial sample,
we obtain another beta density function. For this reason, we say the set of
all beta density functions is a conjugate family of density functions for
binomial sampling.
6.1. LEARNING A SINGLE PARAMETER
309
3.5
3
2.5
2
1.5
1
0.5
0
0.2
0.4 f 0.6
0.8
1
Figure 6.9: The thickly plotted density function is beta(f ; 3, 3) and represents
our prior belief concerning the relative frequency of heads. The thinly plotted
one is beta(f ; 11, 5), and represents our posterior belief after we have seen 8
heads in 10 trials.
Example 6.8 Suppose we have the binomial sample in Example 6.5 and the
data in Example 6.7. Then a = b = 3, s = 8, and t = 2. Due to the preceding
corollary,
ρ(f|d) = beta(f ; 3 + 8, 3 + 2) = beta(f ; 11, 5).
Figure 6.9 shows the original density function and the updated density function.
Suppose we have a sample D of size M , a set d of values (data) of the
variables in D, and we create a sample of size M + 1 by adding another variable
X (M +1) to the sample. Then the conditional probability distribution of X (M+1)
given d is called the updated distribution relative to the sample and the data
d. It represents our posterior belief concerning the next trial.
In the case of binomial samples, E(F |d) is our posterior estimate of the
relative frequency with which X = 1. The following theorem shows that it is
the same as P (X (M +1) = 1|d).
Theorem 6.4 Suppose the conditions in Theorem 6.2 hold, and we create a
binomial sample of size M + 1 by adding another variable X (M +1) . Then if D
is the binomial sample of size M , the updated distribution relative to the sample
and data d is given by
´
³
P X (M +1) = 1|d = E(F |d).
310
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
Proof. We have
Z 1
³
´
P X (M +1) = 1|d
=
P (X (M +1) = 1|f, d)ρ(f|d)df
0
Z 1
=
P (X (M +1) = 1|f )ρ(f |d)df
0
Z 1
=
f ρ(f |d)df
0
= E(F |d).
The second equality above is because X (M +1) is independent of the variables
in D conditional on F , and the third is obtained by applying Equality 6.2 in
Definition 6.3.
Corollary 6.4 If the conditions in Theorem 6.4 hold and F has a beta distribution with parameters a, b, N = a + b, then
´
³
a+s
.
P X (M +1) = 1|d =
N +M
Proof. The proof follows immediately from Theorem 6.4, Corollary 6.3, and
Lemma 6.3.
Notice that our probability (which is also our estimate of the relative frequency) converges to the relative frequency.
Example 6.9 Suppose we have the binomial sample in Example 6.5 and the
data in Example 6.7. Then a = 3, N = 6, s = 8, and M = 10. Due to the
preceding corollary, the probability of heads for the 11th toss is given by
´
³
3+8
= .6875.
P X (M +1) = 1|d =
6 + 10
Furthermore, .6875 is our new estimate of the relative frequency with which the
coin will land heads.
In a Bayesian network representation of our posterior beliefs based on a
sample, we often drop the superscript on X and represent our beliefs as shown
in Figure 6.10. That figure shows our prior beliefs and posterior beliefs in the
case of the previous examples.
6.2
More on the Beta Density Function
First we discuss the case where the parameters a and b in the beta density
function are not integers. Then we present guidelines for assessing the values of
a and b. Finally, we show an argument for using the beta density function to
quantify our beliefs.
6.2. MORE ON THE BETA DENSITY FUNCTION
beta(f; 3,3)
311
beta(f; 11,5)
F
F
X
X
P(X = 1| f) = f
P(X = 1| f) = f
(a)
(b)
Figure 6.10: The Bayesian network in (a) shows our prior belief concerning the
relative frequency of heads, and the one in (b) shows our posterior belief after
we have seen 8 heads in 10 trials.
6.2.1
Non-integral Values of a and b
So far we have only shown examples where a and b are both integers ≥ 1. Next
we discuss nonintegral values. Figure 6.11 shows the beta(f ; .2, .2) density function. As that figure illustrates, as a and b approach 0, we become increasingly
certain the relative frequency is either 0 or 1.
You may wonder when we would ever use non-integral values of a and b.
The following example gives such a case.
Example 6.10 This example is taken from [Berry, 1996]. Glass panels in highrise buildings sometimes break and fall to the ground. A particular case involved
39 broken panels. In their quest for determining why the panels broke, the owners
wanted to analyze the broken panels, but they could only recover three of them.
These three were found to contain Nickel Sulfide (NiS), a manufacturing flaw
which can cause panels to break. In order to determine whether they should hold
the manufacturer responsible, the owners then wanted to determine how probable
it was that all 39 panels contained NiS. So they contacted a glass expert.
The glass expert testified that among glass panels that break, only 5% contain
NiS. However, NiS tends to be pervasive in production lots. So given the first
panel sampled, from a particular production lot of broken panels, contains NiS,
the expert felt the probability is .95 that the second panel sampled also does. It
was known all 39 panels came from the same production lot. Let X (h) s value be
either N iS or qN iS depending on whether the hth panel contains NiS. Given the
above, the expert’s probabilities are as follows (All probabilities are conditional
312
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
5
4
3
2
1
0
0.2
0.4 f 0.6
0.8
1
Figure 6.11: The beta(f ; .2, .2) density function.
0.2
0.18
0.16
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
0.2
0.4 f 0.6
0.8
1
Figure 6.12: The thickly plotted density function represents an expert’s prior
belief concerning the relative frequency of NiS in glass panels, while the thinly
plotted one represents his belief after he learns 3 window panels contain NiS.
6.2. MORE ON THE BETA DENSITY FUNCTION
313
on the panels breaking and coming from the same production lot):
´
³
P X (1) = NiS = .05
³
´
P X (2) = N iS|X (1) = N iS = .95.
So, if we model the expert’s beliefs with a beta density function, Corollaries 6.1
and 6.4 imply
a
= .05
a+b
a+1
= .95.
a+1+b
and
Solving for a and b yields
a=
1
360
b=
19
.
360
This is an alternative technique for assessing a and b. Namely, we assess the
probability for the first trial. Then we assess the conditional probability for the
second trial given the first one is a ‘success’. So the expert’s belief concerning
the relative frequency of NiS is beta(f ; 1/360, 19/360). Therefore, after sampling
3 windows and seeing they all contain NiS, due to Corollary 6.3, his updated
belief is beta(f; 3+1/360, 19/360). The original density function and the updated
one are shown in Figure 6.12. Notice how his belief changes from believing the
relative frequency is either very low or very high (with more density on the low
end) to believing it is most likely very high.
Furthermore, due to Corollary 6.4, we have that the probability any one of
the other 36 panels (the next one sampled) contains NiS is given by
a+s
=
N +M
1
360
20
360
+3
= .983.
+3
We are really most interested in whether all 36 remaining panels contain
NiS. We can determine the probability of this event as follows. The expert’s
current belief, based on the 3 windows sampled, is beta(f ; 3 + 1/360, 19/360).
Therefore, owing to Corollary 6.2, the probability of the next 36 containing NiS
is equal to
¢
¡
¢ ¡ 19
¢
¡
20
1
Γ 3 + 360
+ 36 Γ 360
+0
Γ 3 + 360
¢ ¡ 19 ¢
¡
¢
¡
= .866.
20
1
Γ 360
Γ 3 + 360
+ 36
Γ 3 + 360
6.2.2
Assessing the Values of a and b
Next we give some guidelines for choosing the size of a and b in the beta density
function, when we are accessing our beliefs concerning a relative frequency.
• a = b = 1: These values mean we consider all numbers in [0, 1] equally
likely to be the relative frequency with which the random variable assumes
314
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
each of its values. We would use these values when we feel we have no
knowledge at all concerning the value of the relative frequency. We might
also use these values to try to achieve objectivity in the sense that we
impose none of our beliefs concerning the relative frequency on the learning
algorithm. We only impose the fact that we know at most two things can
happen. An example might be when we are learning the probability of
lung cancer given smoking from data, and we want to communicate our
result to the scientific community. The scientific community would not
be interested in our prior belief, but in what the data had to say. Note
that we might not actually believe a priori that all relative frequencies are
equally probable, but our goal is not to impose this belief on the learning
algorithm. Essentially the posterior probability represents the belief of an
agent that has no prior belief concerning the relative frequency.
• a, b > 1: These values mean we feel it more probable that the relative
frequency with which the random variable assumes its first value is around
a/(a + b). The larger the values of a and b, the more we believe this. We
would use such values when we want to impose our beliefs concerning
the relative frequency on the learning algorithm. For example, if we were
going to toss a coin taken from the pocket, we might take a = b = 50.
• a, b < 1: These values mean we feel it more probable that the relative
frequency with which the random variable assumes one of its values is low,
although we are not committed to which it is. If we take a = b ≈ 0 (almost
0), then we are almost certain the relative frequency with which it assumes
one value is very close to 1. We would also use values like these when
we want to impose our beliefs concerning the relative frequency on the
learning algorithm. Example 6.10 shows a case in which we would choose
values less than 1. Notice that such prior beliefs are quickly overwhelmed
by data. For example, if a = b = .1, and our data d consists of seeing X
take the value 1 in a single trial,
´ .1 + 1
³
= .917.
P X (M +1) = 1|d =
.2 + 1
(6.3)
Intuitively, we thought a priori that the relative frequency with which X
assumes one of its values is high. The fact that it took the value 1 once
makes us believe it is probably that value.
• a < 1, b > 1: These values mean we feel it is more probable the relative
frequency with which X assumes its first value is low. We would also
use values like these when we want to impose our beliefs concerning the
relative frequency on the learning algorithm
You may wonder, why, when we assume no knowledge at all about the relative frequency (a = b = 1), and the data d contains s occurrences of 1 in N
6.2. MORE ON THE BETA DENSITY FUNCTION
315
trials, we have, due to Corollary 6.4, that
³
´
1+s
P X (M +1) = 1|d =
.
2+N
You may feel the answer should be s/N because this is all the data has to say.
An intuition for this is obtained by looking at an example where our sample size
is one. Suppose I am going to sample dogs and determine whether or not they
eat the potato chips which I offer them. Since I have no idea as to whether a
particular dog would eat potato chips, I assign a = b = 1. Suppose further it
turns out the first dog sampled eats the potato chips. Using Equality 6.3, we
obtain a probability of .67, whereas s/N = 1. The first value (.67) seems to
more accurately model my belief for the next trial. In terms of a comparison to
an urn lottery, it means I would be indifferent between receiving a small prize if
the next dog sampled ate the potato chips and receiving the same small prize if
a white ball were drawn from an urn containing 67% white balls. However, the
second value (1) means I would be indifferent between receiving a small prize if
the next dog sampled ate the potato chips and receiving the same small prize
if a white ball were drawn from an urn containing 100% white balls. I would
feel this way only if I were certain the next dog would eat potato chips, and I
would not be certain of this after one trial. Note that if I take a = b ≈ 0 (almost
0), then my updated belief after one trial would be almost 1. This makes sense
since, in this case, my prior belief was that I was almost certain that either all
dogs would eat potato chips or all would not. So after, seeing one eat them, I
become almost certain all will eat them.
6.2.3
Why the Beta Density Function?
The use of the beta distribution to represent our beliefs concerning a relative
frequency is intuitively appealing and has a long tradition. Indeed, in 1889 the
actuary G.F. Hardy and in 1897 the mathematician W.A. Whitworth proposed
quantifying prior beliefs with beta distributions. However, are there any cogent
arguments we should such distributions? Next we present two such arguments.
First, if we initially consider all numbers in [0, 1] equally likely to be the
relative frequency and therefore use the uniform density function to represent
our prior beliefs, it is a mathematical consequence of the theory that the updated
density function is beta.
Second, in 1982 Sandy Zabell proved that, if we make certain assumptions
about an individual’s beliefs, then that individual must use the beta density
function to quantify any prior beliefs about a relative frequency. Zabell’s theorem actually concerns the Dirichlet distribution, which is a generalization of the
beta distribution. So before giving the theorem, we briefly review the Dirichlet
distribution. This distribution is discussed in detail in the next chapter.
Definition 6.4 The Dirichlet density function with parameters a1 , a2 , . . . ar ,
316
N=
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
Pr
k=1
ak , where a1 , a2 , . . . ar are integers ≥ 1, is
ρ(f1 , f2 , . . . fr−1 ) =
Γ(N )
a
f1 1−1 f2a2 −1 · · · frar −1
r
Q
Γ(ak )
0 ≤ fk ≤ 1,
k=1
r
X
fk = 1.
k=1
Random variables F1 , F2 , . . . Fr , that have this density function, are said to have
a Dirichlet distribution.
The Dirichlet density function is denoted Dir(f1 , f2 , . . . fr−1 ; a1 , a2 , . . . ar ).
Note that the value of Fr isPuniquely determined by the values of the first
r−1
r − 1 variables (i.e. fr = 1 − h=1 fh ). That is why ρ is only a function of
r − 1 variables. Let’s show the beta density function is the same as the Dirichlet
density function for r = 2. In this case, the Dirichlet density function is
ρ(f1 ) =
=
Γ(N )
a
f 1−1 f2a2 −1
Γ(a1 )Γ(a2 ) 1
Γ(N )
a
f 1−1 (1 − f1 )a2 −1
Γ(a1 )Γ(a2 ) 1
0 ≤ fk ≤ 1, f1 + f2 = 1
0 ≤ f1 ≤ 1,
which is the beta density function.
As mentioned above, we show examples of Dirichlet density functions and
prove properties of the Dirichlet family in the next chapter. Presently, we only
give Zabell’s theorem. First we need some definitions and discussion. We start
with a formal definition of exchangeability:
Definition 6.5 Let D = {X (1) , X (2) , . . . X M } be an ordered set (sequence) of
random variables, each with space {1, 2, . . . r}. If for every two sets of values
d0 and d00 of the variables in D, such that each of the r values occurs the same
number of times in d0 and d00 , we have
P (d0 ) = P (d00 ),
the sequence is said to be finite exchangeable.
Note that P above represents an individual’s beliefs. So exchangeability is
relative to an individual.
Example 6.11 Suppose D = {X (1) , X (2) , . . . X 7 } represent the outcomes of
seven throws of a six-sided die. Then if we assume the sequence is finite exchangeable,
P ({1, 3, 2, 1, 4, 2, 1}) = P ({1, 1, 1, 2, 2, 3, 4})
because in both sets 1 occurs three times, 2 occurs twice, 3 and 4 each occur
once, and 5 and 6 do not occur at all.
Definition 6.6 Let X (1) , X (2) , X (3) , . . . be an infinite sequence of random variables, each with space {1, 2, . . . r}. If for every M , the sequence of the first M
variables is finite exchangeable, the sequence is said to be infinite exchangeable.
6.2. MORE ON THE BETA DENSITY FUNCTION
317
Exchangeability seems to be a minimal assumption in the case of experiments
which are in some sense considered random processes (Indeed, recall we offered
exchangeability as a Bayesian definition of ‘random process’ in Section 6.1.2.).
For example, it seems many individuals would assign the same probability to
the two sequences of die throws in Example 6.11.
Definition 6.7 Let X (1) , . . . X (M ) , X (M +1) be a sequence of random variables,
each with space {1, 2, . . . r}, and let D = {X (1) , X (2) , . . . X (M ) }. Suppose for
every set of values d of the variables in D, we have
P (X (M +1) = k|d) = gk (sk , M ),
where sk is the number of times k occurs in d. That is, the probability the next
variable is equal to k is a function gk only of how many times k occurs in d and
the number of variables in D. Then Johnson’s sufficientness postulate is
said to hold for the sequence.
Example 6.12 Suppose we throw a six-sided die seven times and we have the
following set of outcomes:
d = {2, 3, 1, 2, 1, 2, 4}.
Since we repeat the experiment seven times, M is equal to seven and since 2
occurs three times, s2 is equal to three. Johnson’s sufficientness postulate says
our probability of the eighth throw landing 2 can be determined using a twoargument function evaluated at three (the value of s2 ) and seven (the value of
M ).
Johnson’s sufficientness hypothesis also seems to be a minimal assumption
in the case of experiments which are in some sense considered to be random
processes. That is, the probability of a given outcome for the next trial can be
computed using the same function for all possible outcomes, and this function
needs only the number of previous trials and the number of times the outcome
occurred. However, given only this hypothesis and the assumption of exchangeability, we get the surprising result that an individual’s prior beliefs concerning
the relative frequency must be represented by a Dirichlet distribution. We give
that theorem after a lemma:
Lemma 6.6 Suppose Johnson’s sufficientness postulate holds for X (1) , . . . X (M ) ,
X (M +1) and the number of values r in their space is greater than 2. Then there
exists constants a1 , a2 , . . . ar ≥ 0 and b such that for every set of values d of the
variables in D = {X (1) , X (2) , . . . X (M ) },
´
³
P X (M +1) = k|d = ak + bsk ,
where sk is the number of times k occurs in d.
Proof. The proof can be found in [Zabell, 1982].
318
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
Theorem 6.5 Suppose X (1) , X (2) , X (3) , . . . is an infinite exchangeable sequence
of random variables, which are not independent, such that for every M , there
exists constants a©1 , a2 , . . . ar ≥ 0 and bªsuch that for every set of values d of the
variables in D = X (1) , X (2) , . . . X (M) ,
³
´
P X (M +1) = k|d = ak + bsk ,
(6.4)
where sk is the number of times k occurs in d (Note that the values of the
constants depend on M even though we do not explicitly denote that.) Let
F1 , F2 , . . . Fr be random variables such that for each value fk of Fk such that
0 ≤ fk ≤ 1
´
³
(6.5)
P X (1) = k|fk = fk .
Then the distribution of the Fk s is Dirichlet.
Proof. The proof can be found in [Zabell, 1982].
Lemma 6.6 and Theorem 6.5 together show that if r > 2, then the assumptions of exchangeability, Johnson’s sufficientness postulate, and Equality
6.5 imply the prior distribution of the relative frequencies must be Dirichlet. If
r = 2, we can conclude the prior distributions must be beta from exchangeability
and the linearity condition in Equality 6.4.
Johnson’s sufficientness postulate seems most plausible. However, there are
situations in which it seems to be too strong of an assumption. For example,
while engaged in cryptanalytic work in for the British government in World War
II, the logician Alan Turing noticed that the frequencies of the frequencies may
contain information relative to the probability of each value. That is, Turing
noticed the following. Let ci be the number of frequencies equal to i. For
example, suppose there are r = 5 values and we observe the sequence
{2, 1, 5, 3, 4, 2, 3, 4, 2, 1, 2}.
Then c1 = 1 because 5 occurs once, c2 = 3 because 1, 3, and 4 each occur twice,
c3 = 0 because no value occurs three times, and c4 = 1 because 2 occurs four
times. Although it is not obvious, the ci ’s can sometimes be used to determine
probabilities of the values (See [Good, 1965].).
The use of the Dirichlet distribution to quantify our prior beliefs concerning
relative frequencies only concerns the case where we know the number of values
of the variable in advance. For example, we know a thumbtack can land two
ways, we know a die can land six ways, we know a patient either does or does
not have lung cancer. In some cases, we know the variable but we do not know
how many values it can have. For example, if I were about to be stranded on a
desert island, I would expect to find some species there, but I would not know
how many different species there might be. After seeing one creature of a given
species, I might want to attach a probability to the event the next creature
seen will be of that same species and one minus that probability to the event
that it will be of a new species. If the next creature turns out to be of a new
6.3. COMPUTING A PROBABILITY INTERVAL
319
species, I would have three possibilities for the third creature, and so on. This
situation, which is not as relevant to Bayesian networks, is discussed in detail
in [Zabell, 1996].
A derivation of the Dirichlet distribution, based on assumptions different
than those of Zabell, can be found in [Geiger and Heckerman, 1997].
6.3
Computing a Probability Interval
Given a distribution (either prior or posterior) of a random variable F , which
represents our belief concerning a relative frequency, Theorem 6.1 tells us the
expected value of F , which is our estimate of the relative frequency, is also our
probability for the next trial. We would not only be interested in this expected
value, but also in how probable it is that the true relative frequency is close to
this expected value. For example, we may want a value c such that
P (f ∈ (E(F ) − c, E(F ) + c)) = .95.
The interval (E(F ) − c, E(F ) + c) is called a 95% probability interval for F .
Probability intervals can be computed directly from the distribution of F as the
following examples illustrate.
Example 6.13 Suppose we have the binomial sample in Example 6.5. Then
due to Lemma 6.3, the prior expected value is given by
E(F ) =
3
= .5.
3+3
A prior 95% probability interval can therefore be found by solving the following
equation for c:
Z .5+c
Γ(6)
f 2 (1 − f )2 df = .95.
Γ(3)Γ(3)
.5−c
Using the mathematics package Maple, we obtain the solution c = .353. So our
95% probability interval is
(.5 − .353, .5 + .353) = (.147, .853).
Note that we are not all that confident the relative frequency really is around .5.
Suppose now we have the data in Example 6.7. Due to Corollary 6.3, the
posterior density function is beta(f; 3 + 8, 3 + 2) = beta(f ; 11, 5). Therefore, our
posterior expected value is given by
E(F |d) =
11
= .688,
11 + 5
and a posterior 95% probability interval can be found by solving
Z .688+c
Γ(16)
f 10 (1 − f )4 df = .95.
Γ(11)Γ(5)
.688−c
320
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
3.5
3
2.5
2
1.5
1
0.5
00
0.2
0.4
f
0.6
0.8
.474
1
.932
Figure 6.13: The beta(f ; 11, 5) density function. The dotted lines show a 95%
probability interval for F . This means 95% of the area under the curve is
between .474 and .932.
Using Maple, we obtain the solution c = .214. So our 95% probability interval
is
(.688 − .214, .688 + .214) = (.474, .902).
Note that the width of our interval has decreased significantly in light of the
data. This interval is shown in Figure 6.13.
Example 6.14 Suppose a = 31 and b = 1. Then
E(F ) =
31
= .969.
31 + 1
If we try to obtain a 95% probability interval by solving
Z
.969+c
.969−c
Γ(32)
f 30 (1 − f )0 df = .95,
Γ(31)Γ(1)
we obtain c = .033 and our 95% probability interval is (.936, 1.002). The reason
we obtain this result is that .969 is too close to 1 for there to be an interval,
centered at .969 and contained in the interval (0, 1), that yields 95% of the area
under the curve. In this case, we should solve
Z
1
.969−c
Γ(32)
f 30 (1 − f )0 df = .95
Γ(31)Γ(1)
6.3. COMPUTING A PROBABILITY INTERVAL
321
to obtain c = .061, which yields a 95% probability interval of
(.969 − .061, 1) = (.908, 1).
We obtain a 95% probability interval that is not centered at .969 but at least has
.969 as close to its center as possible.
In general, the procedure for obtaining a perc % probability interval is as
follows:
Computing a perc % Probability Interval for E(F )
Solve the following equation for c:
Z E(F )+c
ρ(f)df = .perc
.E(F )−c
Case (E(F ) − c, E(F ) + c) ⊆ (0, 1):
A perc% probability interval is given by
(E(F ) − c, E(F ) + c) .
Case (E(F ) − c, E(F ) + c) * (0, 1) and E(F ) > .5:
Solve the following equation for c:
Z 1
ρ(f)df = .perc
.E(F )−c
A perc% probability interval is given by
(E(F ) − c, 1) .
Case (E(F ) − c, E(F ) + c) * (0, 1) and E(F ) < .5:
Solve the following equation for c:
Z E(F )+c
ρ(f)df = .perc
.0
A perc % probability interval is given by
(0, E(F ) + c) .
322
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
Example 6.15 Suppose the density function for F is beta(f; 2, 51), and we
want a 95% probability interval. We have
E(F ) =
Solving
Z
.038+c
.038−c
2
= .038.
2 + 51
Γ(53)
f 1 (1 − f)50 df = .95
Γ(2)Γ(51)
yields c = −.063. Since (.038 − [−.063], .038 + [−.063]) = (.101, −.025) * (0, 1)
and E(F ) < .5, we solve
Z
.038+c
0
Γ(53)
f 1 (1 − f)50 df = .95
Γ(2)Γ(51)
which yields c = .050. A 95% probability interval is therefore given by
(0, .038 + .050) = (0, .088).
As we shall see in Section 6.6, sometimes we do not have a simple expression
for the density function of F , but we do know the expected value and the
variance. In such cases we can obtain a probability interval using the normal
density function to approximate the given density function. We present that
approximation next.
Computing a perc % Probability Interval for E(F )
Using the Normal Approximation
The normal approximation to a perc % probability interval for F is given by
(E(F ) − zperc σ(F ), E(F ) + zpercσ(F )) ,
where σ(F ) is the standard deviation of F , and zperc is the z-score obtained
from the Standard Normal Table. The following are some typical values for
zperc :
perc
80
95
99
zperc
1.28
1.96
2.58
Before giving an example of the above approximation, we present a lemma
that enables us to compute the variance of a random variable that has the beta
distribution.
6.4. LEARNING PARAMETERS IN A BAYESIAN NETWORK
323
Lemma 6.7 Suppose the random variables F has the beta(f ; a, b) density function. Then
µ
¶µ
¶
a+1
a
2
E(F ) =
.
a+b+1
a+b
Proof. The proof is left as an exercise.
Example 6.16 Suppose F has density function beta(f ; 11, 5). Then due to the
preceding lemma,
µ
¶µ
¶
11 + 1
11
33
E(F 2 ) =
= ,
11 + 5 + 1
11 + 5
68
which means the variance of F is given by
33
−
V (F ) = E(F ) − [E(F )] =
68
2
and therefore
σ(F ) =
2
µ
11
16
¶2
= .012638,
p
√
V (F ) = .012638 = .112.
The normal approximation therefore yields the following 95% probability interval:
(.688 − (1.96)(.112), .688 + (1.96)(.112)) = (.468, .908).
Compare this to the exact answer of (.474, .902) obtained in Example 6.13.
The normal density function becomes a better approximation of beta(f ; a, b)
as a and b become larger and as they become closer to being equal. If a and b
are each at least 5, it is usually a reasonable approximation.
6.4
Learning Parameters in a Bayesian Network
Next we extend the theory for learning a single parameter to learning all the
parameters in a Bayesian network. First we motivate the theory by presenting
urn examples in which we have only two variable networks. Then we formally
introduce augmented Bayesian networks, and we show how they can be used to
learn parameters in a Bayesian network.
6.4.1
Urn Examples
Recall Figure 6.1 showed an urn containing 101 coins, each with a different
propensity (relative frequency) for landing heads. The propensities were uniformly distributed between 0 and 1. Suppose now we have two such urns as
shown in Figure 6.14, we sample a coin from each urn, we toss the coin from the
urn labeled X1 , and then we toss the coin from the urn labeled X2 . For the sake
of simplicity assume the distributions are each the uniform continuous distribution rather than a discrete distribution. Let X1 be a random variable whose
324
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
.98
.99
1.00
.98
.99
1.00
.03
.04
.05
.03
.04
.05
.00
.01
.02
.00
.01
.02
X1
X2
Figure 6.14: Each urn contains 101 coins, each with a different propensity for
landing heads.
value is the result of the first toss, and X2 be a random variable whose value is
the result of the second toss. If we let 1 stand for heads and 2 stand for tails,
the Bayesian network in Figure 6.15 (a) represents the probability distribution
associated with this experiment. Recall from Section 6.1.1 that such a Bayesian
network is called an augmented Bayesian network because it includes nodes
(which we shade) representing our beliefs about relative frequencies. The probability distribution of F11 is our belief concerning the relative frequency with
which the first coin lands heads, while the probability distribution of F21 is our
belief concerning the relative frequency with which the second coin lands heads.
It is not hard to see that the Bayesian network, whose specified conditional
distributions are the marginal distributions of X1 and X2 , contains the joint
distribution of X1 and X2 . We say this Bayesian network is embedded in the
augmented Bayeisan network. ‘Embedded Bayesian network’ is defined formally
in Section 6.4.2. Owing to Corollary 6.1, this network is the one Figure 6.15
(b). From that network we have
µ ¶µ ¶
1
1
1
P (X1 = 1, X2 = 1) = P (X1 = 1)P (X2 = 1) =
=
2
2
4
µ ¶µ ¶
1
1
1
P (X1 = 1, X2 = 2) = P (X1 = 1)P (X2 = 2) =
=
2
2
4
µ ¶µ ¶
1
1
1
P (X1 = 2, X2 = 1) = P (X1 = 2)P (X2 = 1) =
=
2
2
4
µ ¶µ ¶
1
1
1
P (X1 = 2, X2 = 2) = P (X1 = 2)P (X2 = 2) =
= .
2
2
4
6.4. LEARNING PARAMETERS IN A BAYESIAN NETWORK
beta(f11; 1,1)
325
beta(f21; 1,1)
F11
F21
X1
X2
P(X1 = 1| f11) = f11
P(X2 = 1| f21) = f21
(a)
X1
X2
P(X1 = 1) = 1/2
P(X2 = 1) = 1/2
(b)
Figure 6.15: A Bayesian network representing the probability distribution concerning the experiment of sampling and tossing coins from the urns in Figure
6.14 is in (a); a Bayesian network containing the marginal distribution of X1
and X2 is in (b).
Note that these probabilities are not the relative frequencies with which the
outcomes will occur (unless we sample the two coins with propensity .5). Rather
they are our beliefs concerning the first outcome. They are also the relative
frequencies with which the outcomes will occur if we repeatedly sample coins
with replacement and tossed each sample pair once.
Suppose now we repeatedly toss the coins we sampled. Our goal is to update
the probability distributions in the augmented Bayesian network (and therefore
the parameters in the embedded network) based on the data obtained from
these tosses. Intuitively, we might expect, we could update the distributions
of F11 and F21 separately using the techniques developed in Section 6.1.2, and
we could compute the probability of data by multiplying the probability of the
data on X1 by the probability of the data on X2 . The theory developed Section
6.4.3 justifies doing this. Presently, we illustrate the technique in an example.
Example 6.17 Suppose we sample coins from the urns in Figure 6.14, we toss
the coins 7 times, and we obtain the data d in Table 6.1. In that table, the ith
row shows the outcome for the ith pair of tosses. Recall 1 stands for heads and
2 for tails. Let
1. s11 be the number of times X1 equals 1;
2. t11 be the number of times X1 equals 2;
326
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
Case
1
2
3
4
5
6
7
X1
1
1
1
1
2
2
2
X2
1
1
1
2
1
1
2
Table 6.1: Data on 6 cases
3. s21 be the number of times X2 equals 1;
4. t21 be the number of times X2 equals 2.
Updating the distributions of X1 and X2 separately, we have due to Corollary
6.3 that
ρ(f11 |d) = beta(f11 ; a11 + s11 , b11 + t11 )
= beta(f11 ; 1 + 4, 1 + 3) = beta(f11 ; 5, 4)
ρ(f21 |d) = beta(f21 ; a21 + s21 , b21 + t21 )
= beta(f21 ; 1 + 5, 1 + 2) = beta(f21 ; 6, 3).
The updated augmented and embedded Bayesian networks appear in Figures 6.16
(a) and (b) respectively. According to the embedded Bayesian network,
µ ¶µ ¶
5
2
10
P (X1 = 1, X2 = 1) = P (X1 = 1)P (X2 = 1) =
=
9
3
27
µ ¶µ ¶
5
1
5
P (X1 = 1, X2 = 2) = P (X1 = 1)P (X2 = 2) =
=
9
3
27
µ ¶µ ¶
4
2
8
P (X1 = 2, X2 = 1) = P (X1 = 2)P (X2 = 1) =
=
9
3
27
µ ¶µ ¶
4
1
4
P (X1 = 2, X2 = 2) = P (X1 = 2)P (X2 = 2) =
= ,
9
3
27
which is the probability given the data d.
Furthermore, assuming we can compute the probability of the data by multiplying the probability of the data on X1 by the probability of the data on X2 ,
owing to Corollary 6.2 we have
ih
i
h
Γ(a11 +s11 )Γ(b11 +t11 )
Γ(a21 +s21 )Γ(b21 +t21 )
Γ(N11 )
Γ(N21 )
P (d) =
Γ(N11 +M11 )
Γ(a11 )Γ(b11 )
Γ(N21 +M21 )
Γ(a21 )Γ(b21 )
ih
i
h
Γ(2) Γ(1+4)Γ(1+3)
Γ(2) Γ(1+5)Γ(1+2)
=
Γ(2+7)
Γ(1)Γ(1)
Γ(2+7)
Γ(1)Γ(1)
= 2. 125 9 × 10−5 .
6.4. LEARNING PARAMETERS IN A BAYESIAN NETWORK
beta(f11; 5,4)
327
beta(f21; 6,3)
F11
F21
X1
X2
P(X1=1| f11) = f11
P(X2=1| f21) = f21
(a)
X1
X2
P(X2=1) = 2/3
P(X1=1) = 5/9
(b)
Figure 6.16: A Bayesian network containing the posterior probability distribution given data obtained by performing the experiment of sampling and tossing
coins from the urns in Figure 6.14 is in (a); a Bayesian network containing the
posterior marginal distribution of X1 and X2 is in (b).
Note that P (d) is the relative frequency with which we will obtain data d when
we repeatedly sample coins with replacement from the urns and toss each sample
pair 7 times.
Suppose now we have three urns, each containing coins with propensities
uniformly distributed between 0 and 1, as shown in Figure 6.17. Again assume
the distributions are each the uniform continuous distribution. Suppose further
we sample a coin from each urn. We then toss the coin from the urn labeled
X1 . If the result is heads (1), we toss the coin from the urn labeled X2 |X1 = 1,
and if the result is tails (2), we toss the coin from the urn labeled X2 |X1 =
2. Let X1 be a random variable whose value is the result of the first toss,
and let X2 be a random variable whose value is the result of the second toss.
The Bayesian network in Figure 6.18 (a) represents the probability distribution
associated with this experiment. The probability distribution of F11 is our
belief concerning the relative frequency with which the first coin lands heads,
the probability distribution of F21 is our belief concerning the relative frequency
with which the second coin lands heads when the first coin lands heads, and the
probability distribution of F22 is our belief concerning the relative frequency
with which the second coin lands heads when the first coin lands tails. Note the
difference between the network in Figure 6.18 (a) and the one in Figure 6.15
(a). In our second experiment, a result of tossing the coin picked from the urn
328
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
.98
.99
1.00
.03
.04
.05
.00
.01
.02
X1
.98
.99
1.00
.98
.99
1.00
.03
.04
.05
.03
.04
.05
.00
.01
.02
.00
.01
.02
X2|X1=1
X2|X1=2
Figure 6.17: Each urn contains 101 coins, each with a different propensity for
landing heads.
labeled X2 |X1 = 1 tells us nothing about the coin picked from the urn labeled
X2 |X1 = 2. So F21 are F22 independent in Figure 6.18 (a). However, in our
first experiment, the same coin is used in the second toss regardless of the value
of X1 . So in this case F21 and F22 are completely dependent (deterministically
related) and are therefore collapsed into one node in Figure 6.15 (a) (It is also
labeled F21 in that figure.).
It is not hard to see that the Bayesian network, whose specified conditional
distributions are the marginal distributions of X1 , of X2 conditional on X1 = 1,
and of X2 conditional on X1 = 2, contains the joint distribution of X1 and X2 .
This network is shown in Figure 6.15 (b). From that network we have
P (X1 = 1, X2 = 1) = P (X2 = 1|X1 = 1)P (X1 = 1) =
µ ¶µ ¶
1
1
1
=
2
2
4
6.4. LEARNING PARAMETERS IN A BAYESIAN NETWORK
beta(f11; 1,1)
beta(f21; 1,1)
F11
F21
X1
P(X1 = 1| f11) = f11
329
beta(f22; 1,1)
F22
X2
P(X2 = 1|X1 = 1,f21) = f21
P(X2 = 1|X1 = 2,f22) = f22
(a)
X1
X2
P(X1 = 1) = 1/2
P(X2 = 1|X1 = 1) = 1/2
P(X2 = 1|X1 = 2) = 1/2
(b)
Figure 6.18: A Bayesian network representing the probability distribution concerning the experiment (as discussed in the text) of sampling and tossing coins
from the urns in Figure 6.17 is in (a); a Bayesian network containing the marginal distribution of X1 and X2 is in (b).
µ ¶µ ¶
1
1
1
=
2
2
4
µ ¶µ ¶
1
1
1
P (X1 = 2, X2 = 1) = P (X2 = 1|X1 = 2)P (X1 = 2) =
=
2
2
4
µ ¶µ ¶
1
1
1
P (X1 = 2, X2 = 2) = P (X2 = 2|X1 = 2)P (X1 = 2) =
= .
2
2
4
P (X1 = 1, X2 = 2) = P (X2 = 2|X1 = 1)P (X1 = 1) =
Suppose now we repeatedly toss the coins we sampled according to the rules
of the experiment discussed above. Our goal again is to update the probability
distributions in the augmented network (and therefore the parameters in the
embedded network) based on the data obtained from these tosses. Again we
might expect, we could update the distributions F11 , F21 , and F22 separately
using the techniques developed in Section 6.1.2, and we could compute the probability of data by multiplying the probability of the data on X1 , the probability
of the data on X2 when X1 = 1, and the probability of the data on X2 when
X1 = 2. The theory developed Section 6.4.3 justifies doing this. Presently, we
illustrate the technique in an example.
Example 6.18 Suppose we sample coins from the urns in Figure 6.17, we toss
the coins 7 times according to the rules of the experiment discussed above, and
we again obtain the data d in Table 6.1. Let
330
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
beta(f11; 5,4)
beta(f21; 4,2)
F11
F21
X1
P(X1 = 1| f11) = f11
beta(f22; 3,2)
F22
X2
P(X2 = 1|X1 = 1,f21) = f21
P(X2 = 1|X1 = 2,f22) = f22
(a)
X1
P(X1 = 1) = 5/9
X2
P(X2 = 1|X1 = 1) = 2/3
P(X2 = 1|X1 = 2) = 3/5
(b)
Figure 6.19: A Bayesian network containing the posterior probability distribution given data obtained by performing the experiment (as discussed in the
text) of sampling and tossing coins from the urns in Figure 6.17 is in (a); a
Bayesian network containing the posterior marginal distribution of X1 and X2
is in (b).
1. s11 be the number of times X1 equals 1;
2. t11 be the number of times X1 equals 2;
3. s21 be the number of times X2 equals 1 when X1 equals 1;
4. t21 be the number of times X2 equals 2 when X1 equals 1;
5. s22 be the number of times X2 equals 1 when X1 equals 2;
6. t22 be the number of times X2 equals 2 when X1 equals 2.
Updating the distributions separately, we have due to Corollary 6.3 that
ρ(f11 |d) = beta(f11 ; a11 + s11 , b11 + t11 )
= beta(f11 ; 1 + 4, 1 + 3) = beta(f11 ; 5, 4)
ρ(f21 |d) = beta(f21 ; a21 + s21 , b21 + t21 )
= beta(f21 ; 1 + 3, 1 + 1) = beta(f21 ; 4, 2)
ρ(f22 |d) = beta(f22 ; a22 + s22 , b22 + t22 )
= beta(f22 ; 1 + 2, 1 + 1) = beta(f22 ; 3, 2).
6.4. LEARNING PARAMETERS IN A BAYESIAN NETWORK
331
The updated augmented and embedded Bayesian networks appear in Figures 6.19
(a) and (b) respectively. According to the embedded Bayesian network,
µ ¶µ ¶
2
5
10
P (X1 = 1, X2 = 1) = P (X2 = 1|X1 = 1)P (X1 = 1) =
=
3
9
27
µ ¶µ ¶
1
5
5
P (X1 = 1, X2 = 2) = P (X2 = 2|X1 = 1)P (X1 = 1) =
=
3
9
27
µ ¶µ ¶
3
4
4
P (X1 = 2, X2 = 1) = P (X2 = 1|X1 = 2)P (X1 = 2) =
=
5
9
15
µ ¶µ ¶
2
4
8
P (X1 = 2, X2 = 2) = P (X2 = 2|X1 = 2)P (X1 = 2) =
= .
5
9
45
Note that this is not the same distribution as obtained in Example 6.17 even
though the prior distributions are the same.
Furthermore, assuming we can compute the probability of the data by multiplying the probability of the data on X1 , the probability of the data on X2 when
X1 = 1, and the probability of the data on X2 when X1 = 2, owing to Corollary
6.2 we have
i
h
Γ(N11 )
Γ(a11 +s11 )Γ(b11 +t11 )
P (d) =
Γ(N11 +M11 )
Γ(a11 )Γ(b11 )
ih
i
h
Γ(a21 +s21 )Γ(b21 +t21 )
Γ(a22 +s22 )Γ(b22 +t22 )
Γ(N221 )
21 )
× Γ(NΓ(N
Γ(a21 )Γ(b21 )
Γ(N22 +M22 )
Γ(a22 )Γ(b22 )
21 +M21 )
ih
ih
i
h
Γ(2) Γ(1+4)Γ(1+3)
Γ(2) Γ(1+3)Γ(1+1)
Γ(2) Γ(1+2)Γ(1+1)
=
Γ(2+7)
Γ(1)Γ(1)
Γ(2+4)
Γ(1)Γ(1)
Γ(2+3)
Γ(1)Γ(1)
= 1. 488 1 × 10−5 .
Note that P (d) is the relative frequency with which we will obtain data d when we
repeatedly sample coins with replacement from the urns and toss them 7 times
according to the rules of the experiment. Note further that it is not the same as
P (d) obtained in Example 6.17.
6.4.2
Augmented Bayesian Networks
Next we formalize the notions introduced in the previous subsection.
Definition 6.8 An augmented Bayesian network (G, F, ρ) is a Bayesian
network determined by the following:
1. A DAG G = (V, E) where V = {X1 , X2 , . . . Xn } and each Xi is a random
variable.
2. For every i, an auxiliary parent variable Fi of Xi and a density function
ρi of Fi . Each Fi is a root and has no edge to any variable except Xi . The
set of all Fi s is denoted by F. That is,
F = F1 ∪ F2 ∪ · · · Fn .
332
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
3. For every i, for all values pai of the parents PAi in V of Xi , and all values
fi of Fi , a probability distribution of Xi conditional on pai and fi .
In general, in an augmented Bayesian network the distributions of the Fi s
need not be continuous. However, since in the ones we consider they are, we
denote them that way. Furthermore, the conditional distribution of the Xi s
may be either continuous or discrete. Still, we denote the joint distribution of
the Xi s by P .
The idea in an augmented Bayesian network is that we know conditional
independencies among a set of variables, and we are able to represent these
conditional independencies using a DAG G. We then want to represent our
beliefs concerning the unknown conditional relative frequencies (parameters)
needed for that DAG. We do this using the nodes in F. Fi is a set of random
variable representing our belief concerning the relative frequencies of the values
of Xi given values of the parents of Xi .
Clearly, an augmented Bayesian network is simply a Bayesian network. It is
only the notation that distinguishes it.
Since the Fi s are all roots in a Bayesian network, they are mutually independent. Therefore, we have Global Parameter Independence:
ρ(f1 , f2 , . . . fn ) = ρ1 (f1 )ρ2 (f2 ) · · · ρn (fn ).
(6.6)
Subscripting both ρ and f creates clutter. So from now on we will just write the
joint distribution in Equality 6.8 as follows:
ρ(f1 )ρ(f2 ) · · · ρ(fn )
It’s clear from the subscript on f which density function each ρ represents.
We have the following theorem:
Theorem 6.6 Let an augmented Bayesian network (G, F, ρ) be given. Then
the marginal distribution P of {X1 , X2 , . . . Xn } constitutes a Bayesian network
with G. We say (G, F, ρ) embeds (G, P ).
Proof. It is left as an exercise to show that, in general, when we marginalize by
summing (integrating) over the values of a set of roots in a Bayesian network,
such that each root in the set has only one child, the marginal distribution of the
remaining variables constitutes a Bayesian network with the subgraph containing
those variables.
After developing an augmented Bayesian network, its embedded network is
the one used to do inference with the variables in V since this latter network
contains our probability distribution of those variables.
The following augmented Bayesian networks are discussed in this chapter:
Definition 6.9 A binomial augmented Bayesian network (G, F, ρ) is an
augmented Bayesian network as follows:
1. For every i, Xi has space {1, 2}.
6.4. LEARNING PARAMETERS IN A BAYESIAN NETWORK
333
2. For every i, there is an ordering [pai1 , pai2 , . . . paiqi ] of all instantiations
of the parents PAi in V of Xi , where qi is the number of different instantiations of these parents. Furthermore, for every i,
Fi = {Fi1 , Fi2 , . . . Fiqi } ,
where each Fij is a root, has no edge to any variable except Xi , and has
density function
0 ≤ fij ≤ 1.
ρij (fij )
3. For every i and j, and all values fi = {fi1 , . . . fij , . . . fiqi } of Fi ,
P (Xi = 1|paij , fi1 , . . . fij , . . . fiqi ) = fij .
(6.7)
If Xi is a root, PAi is empty. In this case, qi = 1 and paij does not appear
in Equality 6.7.
Since the Fij s are all root in a Bayesian network, they are mutually independent. So besides the Global Parameter Independence of the sets Fi , we have
Local Parameter Independence of their members Fij . That is, for 1 ≤ i ≤ n
ρ(fi1 , fi2 , . . . fiqi ) = ρ(fi1 )ρ(fi2 ) · · · ρ(fiqi ).
Global and local independence together imply
ρ(f11 , f12 , . . . fnqn ) = ρ(f11 )ρ(f12 ) · · · ρ(fnqn ).
(6.8)
Note that again to avoid clutter we did not subscript the density functions.
Figure 6.20 (a) shows a binomial augmented Bayesian network. Note that
we shade the nodes in F. Note further that to avoid clutter in this and future
figures we do not show the conditional distributions in the augmented network.
They are all given by Equality 6.7. In the network in Figure 6.20 (a), q1 = 1,
q2 = 2, q3 = 2, and
PA1 = ∅
pa11 = ∅
PA2 = {X1 }
pa21 = {1}
pa22 = {2}
PA3 = {X2 }
pa31 = {1}
pa32 = {2}.
In a binomial augmented Bayesian network, Fij is a random variable whose
probability distribution represents our belief concerning the relative frequency
with which Xi is equal to 1 given that the parents of Xi are in their jth instantiation. For example, in Figure 6.20 (a), the probability distribution of F11
represents our belief concerning the relative frequency with which X1 is equal to
1, the probability distribution of F21 represents our belief concerning the relative
334
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
beta(f11; 8,2)
F11
beta(f21; 2,6)
beta(f22; 1,1)
F21
X1
beta(f31; 2,1)
F22
X2
F31
beta(f32; 3,4)
F32
X3
(a)
P(X1 = 1)= 4/5
X1
P(X2 = 1|X1 = 1) = 1/4
P(X2 = 1|X1 = 2) = 1/2
X2
P(X3 = 1|X2 = 1) = 2/3
P(X3 = 1|X2 = 2) = 3/7
X3
(b)
Figure 6.20: A binomial augmented Bayesian network is in (a), and its embedded
Bayesian network is in (b).
frequency with which X2 is equal to 1 given that X1 = 1, and the probability
distribution of F22 represents our belief concerning the relative frequency with
which X2 is equal to 1 given that X1 = 2. Furthermore, Equality 6.7 is the same
assumption we made in Section 6.1. Namely, if we knew a relative frequency for
certain, then that relative frequency would be our probability.
Given a binomial augmented Bayesian network (G, F, ρ), the following theorem proves that the conditional probabilities in the embedded Bayesian network
(G, P ) are equal to the expected values of the variables in F.
Theorem 6.7 Let a binomial augmented Bayesian network (G, F, ρ) be given.
Then for each i and each j, the ijth conditional distribution in the embedded
Bayesian network (G, P ) is given by
P (Xi = 1|paij ) = E(Fij ).
Proof. We have
(6.9)
6.4. LEARNING PARAMETERS IN A BAYESIAN NETWORK
335
P (Xi = 1|paij )
Z 1
Z 1
=
···
P (Xi = 1|paij , fi1 , . . . fiqi )ρ(fi1 ) · · · ρ(fiqi )dfi1 · · · dfiqi
0
0
Z 1
Z 1
=
···
fij ρ(fi1 ) · · · ρ(fiqi )dfi1 · · · dfiqi
0
=
Z
0
1
fij ρ(fij )dfij
0
= E(Fij ).
The first equality above is due to the law of total probability and the fact that F is
independent of PAi , the second is due to Equality 6.7, and the third is obtained
by integrating over all density functions other than ρ(fij ).
Corollary 6.5 Let a binomial augmented Bayesian network be given. If each
Fij has a beta distribution with parameters aij , bij , Nij = aij + bij , then for each
i and each j the ijth conditional distribution in the embedded network (G, P ) is
given by
aij
.
P (Xi = 1|paij ) =
Nij
Proof. The proof follows directly from Theorem 6.7 and Lemma 6.3.
Example 6.19 Consider the augmented Bayesian network in Figure 6.20 (a).
We have
E(F11 ) =
E(F21 ) =
E(F22 ) =
E(F31 ) =
E(F32 ) =
8
8+2
2
2+6
1
1+1
2
2+1
3
3+4
=
=
=
=
=
4
5
1
4
1
2
2
3
3
.
7
Therefore, that augmented Bayesian network embeds the Bayesian network in
Figure 6.20 (b).
As mentioned previously, the embedded Bayesian network is the one used
to do inference with the variables in V. For example, after developing the binomial augmented Bayesian network in Figure 6.20 (a), we do inference using the
network in Figure 6.20 (b).
Before ending this subsection, we discuss the assumption concerning our beliefs entailed by a binomial augmented Bayesian network. Namely, it is assumed
that the variables, whose probability distributions represent our belief concerning the relative frequencies, are independent. This assumption holds for the
336
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
urn examples presented in Section 6.4.1. In many cases, however, it seems it
does not hold. For example, suppose in a given population, we find out the relative frequency of lung cancer given smoking is .99. This unusually high value
should make us suspect there is some other carcinogenic present in the population, which means we should now believe the relative frequency of lung cancer
given nonsmoking is higher than we previously thought. Nevertheless, even if
the assumption does not hold in a strict sense, we can still make it because,
regardless of our personal beliefs, the learning algorithm in the next section will
converge to the true joint distribution of the relative frequencies as long the
conditional independencies entailed by the DAG are correct. We might say that
our model represents the beliefs of an agent that has no prior knowledge about
the variables other than these conditional independencies. Note that we can create arcs among the Fij s or introduce hidden variables that connect them, and
thereby model our beliefs concerning dependence of these variables. However,
this method will not be discussed in this text.
6.4.3
Learning Using an Augmented Bayesian Network
Next we develop theory which entails we can update probability distributions
as illustrated in Section 6.4.1. Random vectors (defined in Section 5.3.1) are
used in this development.
We start with the following definitions:
Definition 6.10 Suppose we have a sample of size M as follows:
1. We have the random vectors



(1)
(2)
X1
X1



. 
.
X(1) = 
X(2) = 
 .. 
 ..
(1)
(2)
Xn
Xn




···
o
n
D = X(1) , X(2) , . . . X(M )
(h)
such that for every i each Xi
X(M )

(M)
X1


..

=
.


(M)
Xn

has the same space.
2. There is an augmented Bayesian network (G, F, ρ), where G = (V, E), such
that for 1 ≤ h ≤ M,
(h)
{X1 , . . . Xn(h) }
constitutes an instance of V in G resulting in a distinct augmented Bayesian
network.
Then the sample D is called a Bayesian network sample of size M with
parameter (G, F).
Definition 6.11 Suppose we have a Bayesian network sample of size M such
that
6.4. LEARNING PARAMETERS IN A BAYESIAN NETWORK
337
F
X(1)
X(2)
X(M)
(a)
F11
X(1)1
F21
X(1)2
F22
X(2)1
X(2)2
(b)
Figure 6.21: The high-level structure of a Bayesian network sample is given by
the DAG in (a). In that DAG, each node and arc actually represents a set of
nodes and arcs respectively. The detailed structure in the case of a binomial
augmented Bayesian network sample when m = n = 2 is shown in (b).
(h)
1. for every i each Xi
has space {1, 2};
2. its augmented Bayesian network (G, F, ρ) is binomial.
Then D is called a binomial Bayesian network sample of size M with
parameter (G, F).
Note that the network (G, F, ρ) is used as a schema for representing other
augmented Bayesian networks. Note further that in application each X(h) is a
case; that is, it is a random vector whose value is data on one individual sampled.
Finally, note that a Bayesian network sample is itself a big Bayesian network,
and that for each h the subgraph, consisting of the variables in F united with
the set of variables that comprise each X(h) , constitutes an augmented Bayesian
network with ρ. This is illustrated in Figure 6.21. From that figure, we see that
the X(h) s are mutually independent conditional on F because F d-separates all of
338
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
them. A binomial sample (See Definition 6.3.) is a binomial Bayesian network
sample in which G contains only one node.
The idea in a Bayesian network sample is that we know the conditional independencies in the relative frequency distribution of a set of variables, and we are
able to represent these independencies and our beliefs concerning the relative
frequencies of the variables using an augmented Bayesian network. Then we
obtain data consisting of a set of cases (different instantiations of those variables). Our goal is to update our beliefs concerning the relative frequencies from
the data. To that end, we first obtain results that apply to all Bayesian network samples. After that, we obtain results pertaining specifically to binomial
Bayesian network samples.
Results for all Bayesian Network Samples
Lemma 6.8 Suppose
1. D is a Bayesian network sample of size M with parameter (G, F);
2. we have a set of values (data) of the X(h) s as follows:
x(1)

(1)
x1
 . 

=
 .. 
(1)
x1

x(2)

(2)
x1
 . 

=
 .. 
(2)
xn

···
x(M )

(M )
x1
 . 

=
 .. 
(M )
xn

d = {x(1) , x(2) , . . . x(M ) }.
Then
P (d|f1 , . . . fn ) =
M
n Y
Y
i=1 h=1
(h)
P (x(h)
i |pai , fi ),
where pa(h)
contains the values of the parents of Xi in the hth case.
i
6.4. LEARNING PARAMETERS IN A BAYESIAN NETWORK
339
Proof. We have
P (d|f1 , . . . fn ) =
M
Y
h=1
P (x(h) |f1 , . . . fn )
M
Y
P (x(h) , f1 , . . . fn )
ρ(f1 , . . . fn )
h=1

 n
n
Y
Y
(h)
(h)
P (xi |pai , fi )
ρ(fi ) 
M 
Y

 i=1
i=1


=
n


Y

h=1 
ρ(fi )
=
i=1
=
M Y
n
Y
h=1 i=1
=
M
n Y
Y
i=1 h=1
(h)
(h)
P (xi |pai , fi )
(h)
P (x(h)
i |pai , fi ).
The first equality above is because the X(h) s are mutually independent given F,
and the denominator in the third equality is due to global parameter independence.
Lemma 6.9 Suppose we have the conditions in Lemma 6.8. Then for each i,
P (d|fi ) =
M
Y
h=1
(h)
(h)
P (xi |pai , fi )
M
YZ Y
j6=i f
j
h=1
(h)
(h)
P (xj |paj , fj )ρ(fj )dfj .
Proof. We have
P (d|fi ) =
Z
P (d|f1 , . . . fn )
=
fj 6=fi
=
=
M
n Y
Y
j=1 h=1
(h)
(h)
P (xj |paj , fj )
M
Y
(h)
(h)
P (xi |pai , fi )
M
Y
(h)
(h)
P (xi |pai , fi )
h=1
h=1
[ρ(fj )dfj ]
j6=i
fj 6=fi
Z
Y
Y
Z YY
M
j6=i h=1
fj 6=fi
M
YZ Y
j6=i f
j
[ρ(fj )dfj ]
j6=i
h=1
(h)
(h)
P (xj |paj , fj )
(h)
Y
[ρ(fj )dfj ]
j6=i
(h)
P (xj |paj , fj )ρ(fj )dfj .
The first equality above is due to the law of total probability and global parameter
independence, and the second is due to Lemma 6.8.
340
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
Theorem 6.8 Suppose we have the conditions in Lemma 6.8. Then
P (d) =
n
Y
i=1


Z Y
M
(h)
(h)

P (xi |pai , fi )ρ(fi )dfi  .
fi h=1
Proof. We have
P (d) =
Z
···
Z
···
f1
=
P (d|f1 , . . . fn )ρ(f1 ) · · · ρ(fn )df1 · · · dfn
fn
f1
=
Z
n
Y
i=1


Z Y
M
n Y
fn i=1 h=1
Z Y
M
fi h=1
(h)
P (x(h)
i |pai , fi )ρ(f1 ) · · · ρ(fn )df1 · · · dfn

(h)
.
P (x(h)
i |pai , fi )ρ(fi )dfi
The second equality above is due to Lemma 6.8.
Theorem 6.9 (Posterior Global Parameter Independence) Suppose we
have the conditions in Lemma 6.8. Then the Fi s are mutually independent
conditional on D. That is,
ρ(f1 , . . . fn |d) =
n
Y
i=1
ρ(fi |d).
Proof. We have
P (d|f1 , . . . fn )
ρ(fi )
i=1
ρ(f1 , . . . fn |d) =
=
n
Y
P (d)
M
n Y
Y
i=1 h=1
(h)
(h)
P (xi |pai , fi )
P (d)
n
Y
i=1
ρ(fi )
.
(6.10)
The first equality above is due to Bayes’ Theorem and global parameter independence, and the second is due to Lemma 6.8.
We have further that for each i
6.4. LEARNING PARAMETERS IN A BAYESIAN NETWORK
n
Y
i=1
=
ρ(fi |d)
n
Y
P (d|fi )ρ(fi )
i=1
=
n
Y
i=1
P (d)


Z Y
M
M
Y
Y


(h)
(h)
P (x(h)
P (x(h)

i |pai , fi )
j |paj , fj )ρ(fj )dfj  ρ(fi )
h=1
M
n Y
Y
i=1 h=1
=
341
j6=i f
j
h=1
P (d)

n YZ Y
M
n
Y
Y

(h)
(h)
(h)
(h)
P (xi |pai , fi ) 
P (xj |paj , fj )ρ(fj )dfj 
ρ(fi )

i=1 j6=i f
j
i=1
h=1
n
[P (d)]
(h)
n−1
P (x(h)
i |pai , fi ) [P (d)]
=
M
n Y
Y
(h)
P (x(h)
i |pai , fi )
=
M
n Y
Y
i=1 h=1
i=1 h=1
[P (d)]n
P (d)
n
Y
n
Y
i=1
ρ(fi )
.
ρ(fi )
i=1
(6.11)
The first equality above is due to Bayes’ Theorem, the second is due to Lemma
6.9, and the fourth is obtained by rearranging terms and applying Theorem 6.8.
Since Expressions 6.10 and 6.11 are equal, the theorem is proven.
Before giving our final theorem concerning Bayesian network samples, we
have the following definition:
Definition 6.12 Suppose we have the conditions in Lemma 6.8. Then the augmented Bayesian network (G, F, ρ|d) is called the updated augmented Bayesian network relative to the Bayesian network sample and the data d. The network it embeds is called the updated embedded Bayesian network relative
to the Bayesian network sample and the data d.
Note that ρ|d denotes the density function ρ(f1 , . . . fn |d).
Theorem 6.10 Suppose the conditions in Lemma 6.8 hold, and we create a
Bayesian network sample of size M + 1 by including another random vector


(M+1)
X1


..
.
X(M +1) = 
.


(M+1)
Xn
342
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
Then if D is the Bayesian network sample of size M, the updated distribution
+1)
+1)
P (x(M
, . . . x(M
|d)
n
1
is the probability distribution in the updated embedded Bayesian network.
Proof. We have due to that law of total probability that
(M +1)
(M +1)
, . . . xn
|d)
Z
Z
(M+1)
+1)
···
P (x1
, . . . x(M
|f1 , . . . fn , d)ρ(f1 , . . . fn |d)df1 · · · dfn
=
n
f1
fn
Z
Z
(M+1)
+1)
=
···
P (x1
, . . . x(M
|f1 , . . . fn )ρ(f1 , . . . fn |d)df1 · · · dfn .
n
P (x1
f1
fn
The second equality is because X(M+1) is independent of D conditional on F.
This proves the theorem since this last expression is the probability distribution
in the updated embedded Bayesian network.
Due to Theorem 6.10, the updated embedded Bayesian network is the one
used to do inference for the M + 1st case. When doing inference for that case,
we ordinarily do not use the superscript but rather just use the notation Xi .
Furthermore, we do not show the conditioning on D = d. Essentially, we simply
use the updated embedded network as always representing our current belief for
the next case.
Results for Binomial Bayesian Network Samples
Next we apply the previous results to binomial Bayesian network samples.
Lemma 6.10 Suppose
1. D is a binomial Bayesian network sample of size M with parameter (G, F);
2. we have a set of values (data) of



(1)
x1
 . 


x(2) = 
x(1) = 
 .. 

(1)
x1
the X(h) s as follows:


(2)
(M )
x1
x1


.. 
.
···
x(M ) = 
. 
 ..
(2)
(M )
xn
xn
d = {x(1) , x(2) , . . . x(M ) };




3. Mij is the number of x(h) s in which Xi(h) ’s parents are in their jth instan(h)
tiation, and of these Mij cases, sij is the number in which xi is equal to
1 and tij is the number in which it equals 2.
Then
M
Y
h=1
Proof. Let
(h)
P (x(h)
i |pai , fi ) =
qi
Y
j=1
(fij )sij (1 − fij )tij .
6.4. LEARNING PARAMETERS IN A BAYESIAN NETWORK
343
Hij be the set of all indices h such that Xi(h) ’s parents are in their jth instantiation paij .
We have
M
Y
h=1
(h)
M
Y
(h)
P (xi |pai , fi ) =
h=1
qi
(h)
Y Y
=
j=1 h∈Hij
qi
Y
=
j=1
(h)
P (xi |pai , fi1 , . . . fiqi )
P (x(h)
i |paij , fi1 , . . . fiqi )
(fij )sij (1 − fij )tij .
The first equality above is obtained by substituting the members of fi , the second
is obtained by rearranging terms, and the fifth is due to Equality 6.7.
Lemma 6.11 Suppose we have the conditions in Lemma 6.10. Then
P (d|f11 , . . . fnqn ) =
qi
n Y
Y
i=1 j=1
(fij )sij (1 − fij )tij .
Proof. We have
P (d|f11 , . . . fnqn ) = P (d|f1 , . . . fn )
=
=
M
n Y
Y
i=1 h=1
qi
n Y
Y
i=1 j=1
(h)
P (x(h)
i |pai , fi )
(fij )sij (1 − fij )tij .
The first equality above is obtained by replacing the members of the fi s by these
sets, the second is due to Lemma 6.8, and the third is due Lemma 6.10.
Theorem 6.11 Suppose we have the conditions in Lemma 6.10. Then
P (d) =
qi
n Y
Y
i=1 j=1
E(Fij sij [1 − Fij ]tij ).
344
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
beta(f11; 1,1)
beta(f21; 1,1)
beta(f22; 1,1)
F21
F22
F11
X1
beta(f11; 4,6)
beta(f21; 2,3)
beta(f22; 4,3)
F21
F22
F11
X2
X1
(a)
(c)
X1
X1
X2
P(X1 = 1) = 1/2
X2
X2
P(X1 = 1) = 2/5
P(X2 = 1|X1 = 1) = 1/2
P(X2 = 1|X1 = 1) = 2/5
P(X2 = 1|X1 = 2) = 4/7
P(X2 = 1|X1 = 2) = 1/2
(b)
(d)
Figure 6.22: An augmented Bayesian network is in (a) and its embedded
Bayesian network is in (b). Updated networks are in (c) and (d).
Proof. We have
P (d) =
ÃZ M
n
Y
Y
i=1
=
n
Y
i=1
=
=


fi h=1
Z Y
qi
fi j=1
qi Z
n Y
Y
i=1 j=1
qi
n Y
Y
i=1 j=1
0
1
(h)
P (x(h)
i |pai , fi )ρ(fi )dfi
!

(fij )sij (1 − fij )tij ρ(fi )dfi 
(fij )sij (1 − fij )tij ρ(fij )dfij
E(Fij sij [1 − Fij ]tij ).
The first equality is due to Theorem 6.8, and the second is due Lemma 6.10.
Corollary 6.6 Suppose we have the conditions in Lemma 6.10 and each Fij
has a beta distribution with parameters aij , bij , Nij = aij + bij . Then
P (d) =
qi
n Y
Y
Γ(Nij )
Γ(aij + sij )Γ(bij + tij )
.
Γ(N
+
M
)
Γ(aij )Γ(bij )
ij
ij
i=1 j=1
Proof. The proof follows immediately from Theorem 6.11 and Lemma 6.4.
6.4. LEARNING PARAMETERS IN A BAYESIAN NETWORK
Case
1
2
3
4
5
6
7
8
X1
1
1
2
2
2
2
1
2
345
X2
2
1
1
2
1
1
2
2
Table 6.2: Data on 8 cases
Example 6.20 Suppose we have a binomial Bayesian network sample whose
parameter is the augmented Bayesian network in Figure 6.22 (a). For the sake
of concreteness, let’s say the variables represent the following:
Variable
X1
X2
Value
1
2
1
2
When the Variable Takes this Value
There is a history of smoking
There is no history of smoking
Lung Cancer is present
Lung Cancer is absent
Suppose further that we obtain the data (values of X1 and X2 ) on 8 individuals
(cases) shown in Table 6.2. Then
µ ¶
µ ¶
µ ¶
µ ¶
1
1
2
2
(1)
(2)
(3)
(4)
x =
x =
x =
x =
2
1
1
2
µ ¶
µ ¶
µ ¶
µ ¶
2
2
1
2
x(5) =
x(6) =
x(7) =
x(8) =
1
1
2
2
d = {x(1) , x(2) , . . . x(8) }.
Counting yields s11 = 3, t11 = 5, s21 = 1, t21 = 2, s22 = 3, t22 = 2. From
Figure 6.22 (a), we see for all i and j that aij = bij = 1. Therefore, due to the
preceding corollary,
´³
´³
´
³
Γ(2) Γ(1+3)Γ(1+5)
Γ(2) Γ(1+1)Γ(1+2)
Γ(2) Γ(1+3)Γ(1+2)
P (d) =
Γ(2+8)
Γ(1)Γ(1)
Γ(2+3)
Γ(1)Γ(1)
Γ(2+5)
Γ(1)Γ(1)
= 2. 7557 × 10−6 .
Theorem 6.12 (Posterior Local Parameter Independence) Suppose we
have the conditions in Lemma 6.10. Then the Fij s are mutually independent
conditional on D. That is,
ρ(f11 , f12 , . . . fnqn |d) =
qi
n Y
Y
i=1 j=1
ρ(fij |d).
346
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
Furthermore,
ρ(fij |d) =
(fij )sij (1 − fij )tij ρ(fij )
.
E(Fij sij [1 − Fij ]tij )
Proof. We have
ρ(f11 , . . . fnqn |d) =
=
=
=
=
P (d|f11 , . . . fnqn )ρ(f11 ) · · · ρ(fnqn )
P (d)


qi
n Y
Y

(fij )sij (1 − fij )tij  ρ(f11 ) · · · ρ(fnqn )
i=1 j=1
P (d)
qi
n Y
Y
(fij )sij (1 − fij )tij ρ(fij )
qi
n Y
Y
(fij )sij (1 − fij )tij ρ(fij )
i=1 j=1
P (d)
i=1 j=1
qi
n Y
Y
i=1 j=1
qi
n
YY
E(Fij sij [1 − Fij ]tij )
(fij )sij (1 − fij )tij ρ(fij )
.
E(Fij sij [1 − Fij ]tij )
i=1 j=1
The first equality above is due to Bayes’ Theorem, the second is due to Lemma
6.11, the third is obtained by rearranging terms, and the fourth is due to Theorem
6.11.
We have further that for each u and each v
6.4. LEARNING PARAMETERS IN A BAYESIAN NETWORK
ρ(fuv |d) =
=
P (d|fuv )ρ(fuv )
P (d)
!
Ã
R1
R1
Q
· · · 0 P (d|f11 , . . . fnqn )
[ρ(fij )dfij ] ρ(fuv )
0
ij6=uv
suv
(fuv )
=
suv
(fuv )
=
(1 − fuv )
(1 − fuv )
P (d)
Ã
Z
Q
tuv
ij6=uv
tuv
Ã
qi
n Y
Y
i=1 j=1
suv
=
347
Q
ij6=uv
1
sij
(fij )
0
tij
(1 − fij ) ρ(fij )dfij
P (d)
E(Fij
sij
tij
!
ρ(fuv )
!
[1 − Fij ] ) ρ(fuv )
E(Fij sij [1 − Fij ]tij )
(fuv ) (1 − fuv )tuv ρ(fuv )
.
E(Fuv suv [1 − Fuv ]tuv )
The first equality is due to Bayes’ Theorem, the second is due to the law of total
probability and the fact that the Fij s are independent, the third is obtained by
applying Lemma 6.11 and rearranging terms, and the fourth is due to Theorem
6.11.
This proves the theorem.
Corollary 6.7 Suppose we have the conditions in Lemma 6.10 and each Fij
has a beta distribution with parameters aij , bij , Nij = aij + bij . That is, for each
i and each j
ρ(fij ) = beta(fij ; aij , bij ).
Then
ρ(fij |d) = beta(fij ; aij + sij , bij + tij ).
Proof. The proof follows immediately from Theorem 6.12 and Lemma 6.5.
Example 6.21 Suppose we have the binomial Bayesian network sample in Example 6.20. Then since aij = bij = 1 for all i and j, and s11 = 3, t11 = 5, s21
= 1, t21 = 2, s22 = 3, t22 = 2, we have
ρ(f11 |d) = beta(f11 ; 1 + 3, 1 + 5) = beta(f11 ; 4, 6)
ρ(f21 |d) = beta(f21 ; 1 + 1, 1 + 2) = beta(f21 ; 2, 3)
ρ(f22 |d) = beta(f22 ; 1 + 3, 1 + 2) = beta(f22 ; 4, 3).
Recall Definition 6.12 says the augmented Bayesian network (G, F, ρ|d) is
called the updated augmented Bayesian network relative to the Bayesian network sample and the data d, and the network it embeds is called the updated
embedded Bayesian network. The following example shows updated networks.
348
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
Example 6.22 Given the binomial Bayesian network sample in Example 6.20,
the updated networks are the ones shown in Figures 6.22 (c) and (d).
Recall that due to Theorem 6.10, the updated embedded Bayesian network
is the one used to do inference for the M + 1st case, and we do not show the
conditioning on D = d. That is, we simply use the updated embedded network
as always representing our current belief for the next case.
Example 6.23 As discussed in Example 6.22, given the binomial Bayesian network sample in Example 6.20, the updated networks are the ones shown in Figures 6.22 (c) and (d). Therefore, due to Theorem 6.10, the network in Figure
6.22 (d) is the one used to do inference for the 9th case. For example, we
compute P (X2 = 1) for the 9th case as follows:
P (X2 = 1) = P (X2 = 1|X1 = 1)P (X1 = 1) + P (X2 = 1|X1 = 2)P (X1 = 2)
µ ¶µ ¶ µ ¶µ ¶
2
2
4
3
=
+
= .50286.
5
5
7
5
Note in the previous example that we dropped the superscript and the
conditioning on D = d.
6.4.4
A Problem with Updating; Using an Equivalent Sample Size
Let’s compute P (X2 = 1) using the original embedded Bayesian network in
Figure 6.22 (b). We have
µ ¶µ ¶ µ ¶µ ¶
1
1
1
1
P (X2 = 1) =
+
= .5.
2
2
2
2
As shown in Example 6.23, after updating using the data in Table 6.2, we have
P (X2 = 1) = .50286.
(6.12)
Something may seem amiss. We initially had P (X2 = 1) equal to .5. Then we
updated our belief using a sample that had 4 occurrences in which X2 = 1 and
4 occurrences in which X2 = 2, and our P (X2 = 1) changed to .50286. Even
if this seems odd, it is a mathematical consequence of assigning uniform prior
distributions to all three parameters. That is, if the situation being modeled is
the experiment discussed in Section 6.4.1 concerning the three urns in Figure
6.1.1, then this is the correct probability. It is correct in that if we repeated the
experiment of sampling coins and tossing them nine times, the probability in
Equality 6.12 is the relative frequency of the second coin landing heads when the
first eight tosses yield the data in Table 6.2. Although the coin tossing example
clearly illustrates a probability distribution of the value of a relative frequency,
it does not seem to be a good metaphor for applications. Rather it seems
more reasonable to use the metaphor which says our prior belief concerning
6.4. LEARNING PARAMETERS IN A BAYESIAN NETWORK
beta(f11; 1,1)
F11
beta(f12; 1,1)
beta(f21; 1,1)
F12
X1
F21
X2
beta(f11; 2,4)
F11
(b)
beta(f21; 5,5)
X1
F21
X2
(c)
X2
P(X1 = 1|X2 = 1) = 1/2
P(X1 = 1|X2 = 2) = 1/2
beta(f12; 3,3)
F12
(a)
X1
349
P(X2 = 1) = 1/2
X1
X2
P(X1 = 1|X2 = 1) = 1/3
P(X1 = 1|X2 = 2) = 1/2
P(X2 = 1) = 1/2
(d)
Figure 6.23: An augmented Bayesian network is shown in (a) and its embedded
Bayesian network is in (b). Updated networks are shown in (c) and (d).
the relative frequency is obtained from a prior sample. That is, we take the
specified values of aij and bij as meaning our prior experience is equivalent to
having seen a sample in which the first value occurred aij times in aij + bij trials.
Given this, since F11 has the beta(f11 ; 1, 1) density function in Figure 6.22 (a),
our prior experience is equivalent to having seen X1 take the value 1 once in
two trials. However, since F21 also has the beta(f21 ; 1, 1) density function, our
prior experience is equivalent to having seen X2 take the value 1 once out of
the two times X1 took the value 1. Of course, this is not a very reasonable
representation of one’s prior belief. It happened because we have specified four
prior occurrences at node X2 (two in each beta density function), but only two
prior occurrences at node X1 . As a result, we use mixed sample sizes in the
computation of P (X2 = 1). We therefore end up with strange results because
we do not cling to the originally specified probabilities at node X1 as much as
the specifications at node X2 indicate we should.
Another problem arises for the same reason. Suppose we simply reverse
the arc between X1 and X2 , try again to specify prior indifference by using all
beta(f; 1, 1) density functions, and update using the data in Table 6.2. The
results are shown in Figure 6.23. As Figure 6.23 (d) shows, now after updating
we have
P (X2 = 1) = .5.
We see that our updated probabilities depend on which equivalent DAG we
use to represent the independencies. Recall that our assumption in a Bayesian
network sample is that we know the conditional independencies among the variables, and we are merely using the DAG to represent these independencies. So
our results should not depend on which equivalent DAG we use.
350
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
beta(f11; 2,2)
beta(f21; 1,1)
beta(f22; 1,1)
F21
F22
F11
X1
beta(f11; 5,7)
F21
F22
X1
(a)
X2
(c)
X1
X2
P(X1 = 1) = 1/2
beta(f22; 4,3)
F11
X2
X1
beta(f21; 2,3)
P(X2 = 1|X1 = 1) = 1/2
P(X2 = 1|X1 = 2) = 1/2
X2
P(X1 = 1) = 5/12
(b)
P(X2 = 1|X1 = 1) = 2/5
P(X2 = 1|X1 = 2) = 4/7
(d)
Figure 6.24: An augmented Bayesian network is in (a) and its embedded
Bayesian network is in (b). Updated networks are shown in (c) and (d).
Prior Equivalent Sample Size
It seems we could remedy these problems by specifying the same prior sample
size at each node. That is, given the network X1 → X2 , if we specify four
occurrences at X2 using two beta(f ; 1, 1) distributions, then we should specify
four occurrences at X1 using a beta(f ; 2, 2) distribution4 . Figure 6.24 shows the
result when we do this and subsequently update using the data in Table 6.2.
Let’s compute P (X2 = 1) using the updated network in Figure 6.24 (d).
µ ¶µ ¶ µ ¶µ ¶
2
5
4
7
P (X2 = 1) =
+
= .5.
5
12
7
12
So now we get the value we would expect. Furthermore, if we reverse the arrow
between X1 and X2 , specify a beta(f; 2, 2) density function at X2 and two
beta(f; 1, 1) density functions at X1 , and update using the same data, we get
the updated network in Figure 6.25 (d). Clearly, in that network we also have
P (X2 = 1) = .5.
Indeed the entire updated joint distribution for the network with the edge X2 →
X1 is the same as that for the network with the edge X1 → X2 . Example
4 If our prior beliefs are based only on past cases with no missing data, then an equivalent
sample size models our prior beliefs, and this remedy to the noted problems seems appropriate.
However, if our prior beliefs come from different knowledge sources, it may not. For example,
our knowledge of the distribution of X1 may be based on seeing X1 = 1 once in two trials.
However, our knowledge of X2 = 1 given X1 = 1 may be based on a distribution of some
population we read about in a research paper. In this case an equivalent sample size would
not model our prior beliefs. I thank Gregory Cooper for this observation.
6.4. LEARNING PARAMETERS IN A BAYESIAN NETWORK
beta(f11; 1,1)
beta(f21; 2,2)
beta(f21; 1,1)
F11
F12
F21
X1
X2
beta(f11; 2,4)
F11
beta(f21; 3,3)
beta(f21; 6,6)
F12
F21
X1
(a)
X2
(c)
X1
X2
P(X1 = 1|X2 = 1) = 1/2
351
P(X2 = 1) = 1/2
P(X1 = 1|X2 = 2) = 1/2
X1
X2
P(X1 = 1|X2 = 1) = 1/3
P(X2 = 1) = 1/2
P(X1 = 1|X2 = 2) = 1/2
(b)
(d)
Figure 6.25: An augmented Bayesian network is in (a) and its embedded
Bayesian network structure is in (b). Updated networks are shown in (c) and
(d).
6.29, which follows shortly, shows this. So the updated distribution now does
not depend on which equivalent DAG we use. This result is a special case of
Theorem 6.13, which we present shortly. First we need a definition.
Definition 6.13 Suppose we have a binomial augmented Bayesian network in
which the density functions are beta(fij ; aij , bij ) for all i and j. If there is a
number N such that for all i and j
Nij = aij + bij = P (paij ) × N,
(6.13)
then the network is said to have equivalent sample size N .
Recall in the case of a root, PAi is empty and qi = 1. So in this case,
P (pai1 ) = 1. If a binomial augmented Bayesian network has n nodes and equivalent sample size N , we have for 1 ≤ i ≤ n,
qi
X
j=1
Nij =
qi
qi
X
X
£
¤
P (paij ) × N = N ×
P (paij ) = N.
j=1
j=1
The idea in an equivalent sample size is that we specify values of aij and bij
that could actually occur in a sample that exhibit the conditional independencies
entailed by the DAG. Some examples follow:
Example 6.24 Figure 6.26 shows a binomial augmented Bayesian network with
an equivalent sample size of 15. We prove this is the case by showing Equality
352
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
beta(f11; 10,5)
beta(f22; 9,6)
F11
F21
X1
X2
beta(f31; 2,4)
beta(f33; 2,1)
F31
F33
beta(f32; 3,1)
X3
beta(f34; 1,1)
F32
F34
(a)
P(X1 = 1) = 2/3
P(X2 = 1) = 3/5
X1
X2
P(X3 = 1|X1 = 1,X2 = 1) = 1/3
P(X3 = 1|X1 = 1,X2 = 2) = 3/4
X3
P(X3 = 1|X1 = 2,X2 = 1) = 2/3
P(X3 = 1|X1 = 2,X2 = 2) = 1/2
(b)
Figure 6.26: A binomial augmented Bayesian network with an equivalent sample
size of 15 is in (a). It’s embedded network is in (b).
6.13 holds. To that end,
a11 + b11 = 10 + 5 = 15
P (pa11 ) × N = (1) (15) = 15
a31 + b31 = 2 + 4 = 6
µ ¶µ ¶
2
3
P (pa31 ) × N = P (X1 = 1, X2 = 1) × N =
15 = 6.
3
5
It is left as an exercise to compute the remaining 4 pairs of values.
Example 6.25 Figure 6.20 shows a binomial augmented Bayesian network with
an equivalent sample size of 10. We prove this is the case by showing Equality
6.4. LEARNING PARAMETERS IN A BAYESIAN NETWORK
353
6.13 holds. To that end,
a31 + b31 = 2 + 1 = 3
P (pa31 ) × N
= P (X2 = 1) × N
= [P (X2 = 1|X1 = 1)P (X1 = 1) + P (X2 = 1|X1 = 2)P (X1 = 2)] × N
·µ ¶ µ ¶ µ ¶ µ ¶¸
1
4
1
1
=
+
× 10 = 3.
4
5
2
5
It is left as an exercise to compute the remaining 4 pairs of values.
It is unlikely we would arrive at a network with an equivalent sample size
simply by making up values of aij and bij . The next two theorems give common
way for constructing one.
Theorem 6.13 Suppose we specify G, F, and N and assign for all i and j
aij = bij =
N
.
2qi
Then the resultant augmented Bayesian network has equivalent sample size N ,
and the probability distribution in the resultant embedded Bayesian network is
uniform.
Proof. It is left as an exercise to show that with these specifications P (paij ) =
1/qi for all values of i and j. Therefore,
µ ¶
N
N
1
+
=
aij + bij =
× N = P (paij ) × N.
2qi
2qi
qi
It is also left as an exercise to show the probability distribution in the resultant
embedded Bayesian network is uniform. This proves the theorem.
Example 6.26 The specifications in the previous theorem simply spreads the
value of N evenly among all the values specified at a node. This is how the
networks in Figures 6.24 (a) and 6.25 (a) were developed.
Theorem 6.14 Suppose we specify G, F, N , a Bayesian network (G, P ), and
assign for all i and j
aij = P (Xi = 1|paij ) × P (paij ) × N
bij = P (Xi = 2|paij ) × P (paij ) × N.
Then the resultant augmented Bayesian network has equivalent sample size N .
Furthermore, it embeds the originally specified Bayesian network.
Proof. Let P 0 be the probability distribution in the resultant embedded network.
Clearly
aij + bij = P (paij ) × N.
354
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
So if we can show P 0 is the same distribution as P , Equality 6.13 is satisfied
and we are done. To that end, due to Corollary 6.5, we have
P 0 (Xi = 1|paij ) =
=
aij
aij + bij
P (Xi = 1|paij ) × P (paij ) × N
P (paij ) × N
= P (Xi = 1|paij ).
This proves the theorem.
Example 6.27 If we specify the Bayesian network in Figure 6.26 (b), N =
15, and values of aij and bij according to the previous theorem, we obtain the
augmented network in Figure 6.26 (a). It is left as an exercise to do this.
After a definition, a lemma and an example, we prove the theorem to which
we alluded earlier. In what follows we need to refer to two augmented Bayesian
networks. So we show the dependence of F and ρ on G by representing an
augmented Bayesian network as (G, F(G) , ρ|G). The notation ρ|G denotes the
density function in the augmented Bayesian network containing the DAG G. It
does not entail that the DAG G is an event.
Definition 6.14 Binomial augmented Bayesian networks (G1 , F(G1 ) , ρ|G1 ) and
(G2 , F(G2 ) , ρ|G2 ) are called equivalent if they satisfy the following:
1. G1 and G2 are Markov equivalent.
2. The probability distributions in their embedded Bayesian networks are the
same.
3. The specified density functions in both are beta.
4. They have the same equivalent sample size.
Lemma 6.12 (Likelihood Equivalence) Suppose we have two equivalent binomial augmented Bayesian network (G1 , F(G1 ) , ρ|G1 ) and (G2 , F(G2 ) , ρ|G2 ). Let
D be a set of random vectors as specified in Definition 6.11. Then for every set
d of values of the vectors in D,
P (d|G1 ) = P (d|G2 )
where P (d|G1 ) and P (d|G2 ) are the probabilities of d when D is considered a
binomial Bayesian network sample with parameters (G1 , F(G1 ) ) and (G2 , F(G2 ) )
respectively.
Proof. The proof can be found in [Heckerman et al, 1995].
6.4. LEARNING PARAMETERS IN A BAYESIAN NETWORK
355
Example 6.28 Let (G1 , F(G1 ) , ρ|G1 ) be the augmented Bayesian networks in
Figures 6.24 (a) and (G2 , F(G2 ) , ρ|G2 ) be the one in Figure 6.25 (a). Clearly
they are equivalent. Given the data d in Table 6.2, we have due to Corollary 6.2
´³
´³
´
³
Γ(4) Γ(2+3)Γ(2+5)
Γ(2) Γ(1+1)Γ(1+2)
Γ(2) Γ(1+3)Γ(1+2)
P (d|G1 ) =
Γ(4+8)
Γ(2)Γ(2)
Γ(2+3)
Γ(1)Γ(1)
Γ(2+5)
Γ(1)Γ(1)
= 3. 6075 × 10−6 .
P (d|G2 ) =
³
Γ(4) Γ(2+4)Γ(2+4)
Γ(4+8)
Γ(2)Γ(2)
= 3. 6075 × 10−6 .
´³
Γ(2) Γ(1+2)Γ(1+2)
Γ(2+4)
Γ(1)Γ(1)
´³
Γ(2) Γ(1+3)Γ(1+1)
Γ(2+4)
Γ(1)Γ(1)
´
The values are the same as the lemma implies. It is left as an exercise to show
they are not the same for the networks in Figures 6.22 (a) and 6.23 (a).
Theorem 6.15 Suppose we have two equivalent binomial augmented Bayesian
networks (G1 , F(G1 ) , ρ|G1 ) and (G2 , F(G2 ) , ρ|G2 ). Let D be a set of random vectors as specified in Definition 6.11. Then given any set d of values of the vectors
in D, the updated embedded Bayesian network relative to D and the data d, obtained by considering D a binomial Bayesian network sample with parameter
(G1 , F(G1 ) ), contains the same probability distribution as the one obtained by
considering D a binomial Bayesian network sample with parameter (G2 , F(G2 ) ).
Proof. We have
P (x(M+1) |d, G1 ) =
P (d, x(M +1) |G1 )
P (d|G1 )
=
P (d, x(M +1) |G2 )
P (d|G2 )
= P (x(M +1) |d, G2 ).
The second equality is because Lemma 6.12 implies the values in the numerators
and denominators are the same. Since Theorem 6.10 says P (x(M +1) |d, Gi ) is
the probability distribution contained in the updated embedded Bayesian network
relative to D and the data d, the theorem is proven.
Corollary 6.8 Suppose we have two equivalent binomial augmented Bayesian
networks (G1 , F(G1 ) , ρ|G1 ) and (G2 , F(G2 ) , ρ|G2 ). Then given any set d of values of the vectors in D, the updated embedded Bayesian network relative to D
and the data d, obtained by considering D a binomial Bayesian network sample
with parameter (G1 , F(G1 ) ), is equivalent to the one obtained by considering D a
binomial Bayesian network sample with parameter (G2 , F(G2 ) ).
Proof. The proof follows easily from the preceding theorem.
Example 6.29 Consider augmented Bayesian networks discussed in Example
6.28 and their updated embedded Bayesian networks which appear in Figures
6.24 (d) and 6.25 (d). For the one in Figure 6.24 (d), we have
µ ¶µ ¶
2
5
P (X1 = 1, X2 = 1) =
= .16667.
5
12
356
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
For the one in Figure 6.25 (d), we have
µ ¶µ ¶
1
1
P (X1 = 1, X2 = 1) =
= . 16667.
3
2
The values are the same as the theorem implies. It is left as an exercise to check
the other 3 values.
Due to the preceding theorem, as long as we use an equivalent sample size,
our updated probability distribution does not depend on which equivalent DAG
we use to represent a set of conditional independencies. So henceforth we will
always use an equivalent sample size.
Expressing Prior Indifference with a Prior Equivalent Sample Size
Recall from Section 6.2.2 that we take a = b = 1 when we feel all numbers in
[0, 1] equally likely to be the relative frequency with which the random variable
assumes each of its values. We use these values when we feel we have no knowledge at all concerning the value of the relative frequency, and also to try to
achieve objectivity in the sense that we impose none of our beliefs concerning
the relative frequency on the learning algorithm. We tried doing this with the
specifications in the augmented Bayesian network in Figure 6.22 (a), and ended
up with unacceptable results. We eliminated these unacceptable results by using the network with an equivalent sample size in Figure 6.24 (a). However,
in that network, we no longer assign equal density to all possible values of the
relative frequency with which X1 = 1. By the mere fact of including X1 in a
network with X2 , we have become more confident that the relative frequency
with which X1 equals 1 is around .5! So what equivalent sample size should we
use to express prior indifference? We can shed light on this question by looking
again at the two equivalent DAGs X1 → X2 and X2 → X1 . If we used the
former DAG, we would want to specify a beta(f11 ; 1, 1) density function for F11 .
That is, we would want to use a prior sample size of two at node X1 . If we
merely reverse the arrows, it seems there is no reason we should change that
sample size. So it seems we should still use a prior sample size of two at X1 ,
only now we must spread it over two density functions, namely beta(f11 ; .5, .5)
and beta(f12 ; .5, .5). In general, it seems a good way to express prior indifference
is to use an equivalent sample size of two, and for each node to distribute the
sample evenly among all specified values. In this way, the total ‘sample’ size
at each node is always two, even though for non-roots the specified ones are
fractional. Figure 6.27 shows an example.
Our Beliefs When Using a Prior Equivalent Sample Size
At the end of Section 6.4.2, we mentioned that the assumption when using
an augmented Bayesian network is that the variables representing our belief
concerning a relative frequencies are independent. However, when we use different equivalent binomial augmented Bayesian network, different variables are
6.5. LEARNING WITH MISSING DATA ITEMS
beta(f11; 1,1)
357
beta(f22; 1,1)
F11
F21
X1
X2
beta(f31 ; .25,.25)
beta(f33 ; .25,.25)
F31
F33
beta(f32 ; .25,.25)
F32
X3
beta(f34 ; .25,.25)
F34
Figure 6.27: We express prior indifference to the values of all relative frequencies
using a prior equivalent sample size of two.
assumed to be independent. For example, in Figure 6.24, it is assumed that the
variable representing our belief concerning the relative frequency with which
X2 = 1 given X1 = 1 is independent of the variable representing our belief
concerning the relative frequency with which X2 = 1 given X1 = 2. However,
in Figure 6.25, it is assumed that the variable whose probability distribution
represents our belief concerning the relative frequency with which X1 = 1 given
X2 = 1 is independent of the variable whose probability distribution represents
our belief concerning the relative frequency with which X1 = 1 given X2 = 2.
As we have seen, all our results are the same as long as we use equivalent binomial Bayesian networks. So perhaps our assumptions should be stated relative
to using equivalent augmented Bayesian networks rather than a particular one.
We could state this as follows: Given a repeatable experiment whose outcome
determines the state of n random variables, the assumption, when using a binomial augmented Bayesian network, is that our belief, concerning the probability
of the outcome of repeated executions of the experiment, is entailed by any
augmented Bayesian network equivalent to the one used.
6.5
Learning with Missing Data Items
So far we have considered data sets in which every value of every variable is
recorded in every case. Next we consider the case where some data items might
be omitted. How might they be omitted? A common way, and indeed a way
358
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
beta(f11; 2,2)
beta(f21; 1,1)
beta(f22; 1,1)
F21
F22
F11
X1
beta(f11; 6,3)
beta(f21; 7/2,5/2)
F11
X2
F21
X1
(a)
X1
F22
X2
(c)
X1
X2
P(X1 = 1) = 1/2
beta(f22; 3/2,3/2)
X2
P(X1 = 1) = 2/3
P(X2 = 1|X1 = 1) = 1/2
P(X2 = 1|X1 = 1) = 7/12
P(X2 = 1|X1 = 2) = 1/2
P(X2 = 1|X1 = 2) = 1/2
(b)
(d)
Figure 6.28: The network in (a) has been updated to the network in (c) using
a first pass of the EM Algorithm.
Case
1
2
3
4
5
X1
1
1
1
1
2
X2
1
1
1
2
2
Table 6.3: Data on 5 cases
that is relatively easy to handle, is that they are simply random omissions due
to recording problems or some similar error. First we discuss this case.
6.5.1
Data Items Missing at Random
Suppose data items are missing at random. Before discussing how to update
based on such data, let’s review how we update when no data items are missing.
Suppose we want to update the network in Figure 6.28 (a) with the data d in
Table 6.3. Recall that s21 is the number of cases that have X1 equal to 1 and
X2 equal to 1, while t21 is the number of cases that have X1 equal to 1 and X2
equal to 2. So we have
s21
t21
= 3
= 1.
6.5. LEARNING WITH MISSING DATA ITEMS
Case
1
2
3
4
5
X1
1
1
1
1
2
359
X2
1
?
1
2
?
Table 6.4: Data on 5 cases with some data items missing
Case
1
2
2
3
4
5
5
X1
1
1
1
1
1
2
2
# Occurences
1
1/2
1/2
1
1
1/2
1/2
X2
1
1
2
1
2
1
2
Table 6.5: Estimates of missing values
Owing to Corollary 6.7,
ρ(f21 |d) = beta(f21 ; a21 + s21 , b21 + t21 )
= beta(f21 ; 1 + 3, 1 + 1)
= beta(f21 ; 4, 2).
Suppose next that we want to update the network in Figure 6.28 (a) with
the data d in Table 6.4. These data contain missing data items. We do not know
the value of X2 for cases 2 and 5. It seems reasonable to ‘estimate’ the value
of X2 in these cases using P (X2 = 1|X1 = 1). That is, since this probability
equals 1/2, we say X2 has a 1/2 occurrence of 1 in each of cases 2 and 5. So
we replace the data d in Table 6.4 by the data d0 in Table 6.5. We then update
our density functions using the number of occurrences listed in Table 6.5. So
we have
s021
t021
= 1 + 12 + 1 =
= 12 + 1 = 32 ,
s022
t022
=
=
5
2
1
2
1
2,
(6.14)
(6.15)
where s021 , t021 , s022 , and t022 denote the counts in data d0 (shown in Table 6.5).
We then have
ρ(f21 |d0 ) = beta(f21 ; a21 + s021 , b21 + t021 )
¢
¡
= beta f21 ; 1 + 52 , 1 + 32
¢
¡
= beta f21 ; 72 , 52 .
360
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
and
ρ(f22 |d0 ) = beta(f21 ; a22 + s022 , b22 + t022 )
¢
¡
= beta f21 ; 1 + 12 , 1 + 12
¢
¡
= beta f21 ; 32 , 32 .
The updated network is shown in Figure 6.28 (c).
If we consider sij and tij random variables, the method just outlined estimates the actual values of sij and tij by their expected values relative to the
joint distribution of X1 and X2 conditional on the data and on the variables in
F having their prior expected values. That is, if we set
0
0
0
f 0 = {f11
, f21
, f22
} = {f11 , f21 , f22 } = {1/2, 1/2, 1/2}
(The reason for defining f 0 will become clear shortly.), then
s021 = E(s21 |d, f 0 ) =
=
=
5
X
h=1
5
X
h=1
5
X
1 × P (X1(h) = 1, X2(h) = 1|d, f 0 )
(h)
P (X1
(h)
= 1, X2
(6.16)
= 1|x(h) , f 0 )
(h) 0
P (X1(h) = 1, X2(h) = 1|x(h)
1 , x2 , f )
h=1
= 1+
1
2
+ 1 + 0 + 0 = 52 .
Similarly,
t021 = E(t21 |d, f 0 ) = 0 +
1
2
+ 0 + 1 + 0 = 32 .
Furthermore,
s022 = E(s22 |d, f 0 ) =
=
=
5
X
h=1
5
X
h=1
5
X
1 × P (X1(h) = 1, X2(h) = 2|d, f 0 )
(h)
= 1, X2
(h)
= 1, X2
P (X1
P (X1
(h)
= 2|x(h) , f 0 )
(h)
= 2|x1 , x2 , f 0 )
(h)
(h)
h=1
= 0+0+0+0+
1
2
=
1
2
and
t022 = E(t22 |d, f 0 ) = 0 + 0 + 0 + 0 +
1
2
= 12 .
Note that these are the same values obtained in Equalities 6.14 and 6.15.
Using these expected values to estimate our density functions seems reasonable. However, note that our estimates are based only on the ‘data’ in our prior
6.5. LEARNING WITH MISSING DATA ITEMS
361
sample. They are not based on the data d. That is, we say X2 has a 1/2 occurrence of 1 in each of cases 2 and 5 because P (X2 = 1|X1 = 1) = 1/2 according
to our prior sample. However, the data d ‘prefers’ the event X1 = 1, X2 = 1 to
the event X1 = 1, X2 = 2 because the former event occurs twice while the latter
event occurs only once. To incorporate the data d in our estimates we can now
repeat the computation in Expression 6.16 using the probability distribution in
the updated network (Figures 6.28 (c) and (d)). That is, we now set
0
0
0
, f21
, f22
} = {2/3, 7/12, 1/2}
f 0 = {f11
and compute
s021 = E(s21 |d, f 0 ) =
=
5
X
h=1
5
X
(h)
1 × P (X1
(h)
= 1, X2
= 1|d, f 0 )
P (X1(h) = 1, X2(h) = 1|x(h) , f 0 )
h=1
=
5
X
(h)
P (X1
(h)
= 1, X2
(h)
(h)
= 1|x1 , x2 , f 0 )
h=1
= 1+
7
12
7
+ 1 + 0 + 0 = 2 12
.
Similarly,
t021 = E(t21 |d, f 0 ) = 0 +
5
12
5
+ 0 + 1 + 0 = 1 12
.
We re-compute s022 and t022 in the same manner.
Clearly, we can keep repeating the previous two steps. Suppose we reiterate
(v)
(v)
these steps, let sij and tij be the values of s0ij and t0ij after the vth iteration,
and take the
(v)
aij + sij
0
.
(6.17)
lim f = lim
v→∞ ij
v→∞ a + s(v) + b + t(v)
ij
ij
ij
ij
Then under certain regularity conditions, the limit which is approached by
0
0
, . . . fij0 , . . . fnq
} is a value of f that locally maximizes ρ(f|d)5 .
f 0 = {fi1
n
The procedure we have just described is an application of the EM Algorithm
([Dempster et al, 1977], [McLachlan and Krishnan, 1997]). In this algorithm,
the step in which we recompute s0ij and t0ij is called the expectation step,
and the step in which we recompute the value of f 0 is called the maximization
step because we are approaching a local maximum.
The value f̃ which maximizes ρ(f|d) is called the maximum a posterior
probability (MAP) value of f. We want to arrive at this value rather than at
a local maximum. After presenting the algorithm, we discuss a way to avoid a
local maximum.
5 The maximizing values actually depend on the coordinate systems used to express the parameters. The ones given here correspond to the canonical coordinate system for the multinomial distribution (See e.g. [Bernardo and Smith, 1994].).
362
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
Algorithm 6.1 EM-MAP-determination
Problem: Given a binomial augmented Bayesian network in which the density
functions are beta, and data d containing some incomplete data items,
estimate ρ(f|d) and the MAP value of the parameter set f.
Inputs: binomial augmented Bayesian network (G, F, ρ) and data d containing
some incomplete data items.
Outputs: estimate ρ(f|d0 ) of ρ(f|d) and estimate f 0 of the MAP value of the
parameter set f.
void EM _M AP
(augmented-Bayesian-network (G, F, ρ),
data d,
int k,
// number of
// iterations
density-fuction& ρ(f|d0 ),
MAP-estimate& f 0 )
{
float s0ij , t0ij ;
}
for (i = 1; i <= n; i + +)
for (j = 1; j <= qi ; j + +)
assign fij0 a value in the interval (0, 1);
repeat (k times) {
for (i = 1; i <= n; i + +)
// expectation
// step
for (j = 1; j <= qi ; j + +) {
P
(h)
s0ij = E(sij |d, f 0 ) = M
= 1, paij |x(h) , f 0 );
h=1 P (Xi
P
(h)
M
t0ij = E(tij |d, f 0 = h=1 P (Xi = 2, paij |x(h) , f 0 );
}
for (i = 1; i <= n; i + +)
// maximiza// tion step
for (j = 1; j <= qi ; j + +)
aij + s0ij
0
fij
=
;
aij + s0ij + bij + t0ij
}
ρ(fij |d0 ) = beta(fij ; aij + s0ij , bij + t0ij );
Note that in the algorithm we initialized the algorithm by saying “assign fij0
a value in the interval (0, 1)” rather than setting fij0 = aij / (aij + bij ) as we did
in our illustration. We want to end up with the MAP value f̃ of f; however in
general we could end up with a local maximum when starting with any particular
configuration of f 0 . So we do not start at only one particular configuration.
6.5. LEARNING WITH MISSING DATA ITEMS
363
Rather we use multiple restarts of the algorithm. The following is a multiplerestart strategy discussed in [Chickering and Heckerman, 1997]. We sample 64
prior configurations of the variables in F according to a uniform distribution. By
a configuration of the variables we mean an assignment of values to the variables.
Next we perform one expectation and one maximization step, and we retain the
32 initial configurations that yielded the 32 values of f 0 with the largest values
of ρ(f 0 |d). Then we perform two expectation and maximization steps, and we
retain 16 initial configurations using this same rule. We continue in this manner,
in each iteration doubling the number of expectation-maximization steps, until
only one configuration remains. You may wonder how we could determine which
values of f 0 had the largest values of ρ(f 0 |d) when we do not know this density
function. For any value of f we have
ρ(f|d) = αρ(d|f)ρ(f),
which means we can determine whether ρ(f 0 |d) or ρ(f 00 |d) is larger by comparing ρ(d|f 0 )ρ(f 0 ) and ρ(d|f 00 )ρ(f 00 ). To compute ρ(d|f)ρ(f), we simply calculate
Q
(h)
ρ(f) and determine ρ(d|f) = M
|f) using a Bayesian network inference
h=1 P (x
algorithm.
The maximum likelihood (ML) value f̂ of f is the value such that P (d|f)
is a maximum (Recall we introduced this value in Section 4.2.1.). Algorithm
6.1 can be modified to produce the ML value. We simply update as follows:
fij0 =
s0ij
.
s0ij + t0ij
A parameterized EM algorithm, which has faster convergence, is discussed
in [Bauer et al, 1997]. The EM Algorithm is not the only method for handling
missing data items. Other methods include Monte Carlo techniques, in particular Gibb’s sampling, which is discussed in Section 8.3.1.
6.5.2
Data Items Missing Not at Random
The method we outlined in the previous subsection is only appropriate when the
absence of a data item does not depend on the states (values) of other variables.
This is true if the data items are missing at random. It is also true if a variable
is never observed in any cases. However, there are situations where missing data
is not independent of state. For example, in a drug study a patient may become
too sick, due to a side effect of the drug, to complete the study. So the fact that
the result variable is missing depends directly on the value of the side effect variable. [Cooper, 1995a], [Spirtes et al, 1995], and [Ramoni and Sebastiani, 1999]
discuss handling this more complicated situation.
364
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
beta(f11; 4,3)
F11
beta(f21; 3,1)
F21
X
beta(f22; 2,1)
F22
Y
(a)
beta(f11; 3,2)
F11
beta(f12; 1,1)
beta(f21; 5,2)
F12
X
F21
Y
(b)
2
Figure 6.29: E(Fy1
) = 3/10 regardless of which network we use to compute it.
6.6
Variances in Computed Relative Frequencies
Next we discuss how to compute the uncertainty (variance) in a relative frequency for which we have not directly assessed a belief. Rather the relative
frequency estimate is computed from ones for which we have assessed beliefs.
For the sake of space and notational simplicity, in this section we will again
represent variables by unsubscripted letters like X, Y , and Z, and values of
those variables by small letters. For example, the values of X will be x1 and
x2.
6.6.1
A Simple Variance Determination
Consider the Bayesian network in Figure 6.29 (a). The probability distribution of the random variable F11 represents our belief concerning the relative
frequency with which x1 occurs, the probability distribution of F21 represents
our belief concerning the relative frequency with which Y takes the value y1
given that X = x1, and the probability distribution of F22 represents our belief
concerning the relative frequency with which Y takes the value y1 given that
X = x2. Consider now the space determined by the joint distribution of the
Fij s. We assume our belief concerning the relative frequency with which Y takes
6.6. VARIANCES IN COMPUTED RELATIVE FREQUENCIES
365
the value y1 is represented by the random variable Fy1 which assigns
P (y1|f)
to each set of values f = {f11 , f21 , f22 } of the variables in F = {F11 , F21 , F22 },
and our estimate of the relative frequency is E(Fy1 ). Note that this is consistent
with Equality 6.1 in Section 6.1.1. We have
P (y1|f) = P (y1|x1, f)P (x1|f) + P (y1|x2, f)P (x2|f)
= f21 f11 + f22 (1 − f11 ).
The second equality is due to Equality 6.7. Therefore,
Fy1 = F21 F11 + F22 (1 − F11 ).
(6.18)
In Section 6.3 we showed how to compute probability intervals for the Fij s,
but how can we obtain such an interval for Fy1 ? Next we prove two theorems
which enable us to compute the variance V (Fy1 ), from which we can at least
approximate such an interval using a normal approximation.
Theorem 6.16 Suppose F has the beta(f ; a, b) density function. Then
a
a+b
µ
¶µ
¶
a+1
a
2
E(F ) =
a+b+1
a+b
E(F ) =
E(F [1 − F ]) =
ab
.
(a + b + 1) (a + b)
Proof. The proof is left as an exercise.
Theorem 6.17 Suppose the random variable Fy1 is defined as in Expression
6.18, and the Fij s are mutually independent. Then
E(Fy1 ) = E(F21 )E(F11 ) + E(F22 )E(1 − F11 )
2
2
2
) = E(F21
)E(F11
) + 2E(F21 )E(F22 )E(F11 [1 − F11 ])
E(Fy1
2
+E(F22
)E([1 − F11 ]2 ).
Proof. The proof is left as an exercise.
.
Example 6.30 Let’s now compute V (Fy1 ) for the network in Figure 6.29. Due
to Theorem 6.16,
E(F11 ) = E(1 − F11 ) = E(F21 ) = E(F22 ) =
1
1
=
1+1
2
366
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
µ
¶µ
¶
2+1
2
3
=
2+2+1
2+2
10
¶µ
¶
µ
1
1
1+1
2
2
E(F21
=
) = E(F22
)=
1+1+1
1+1
3
2
) = E([1 − F11 ]2 ) =
E(F11
E(F11 [1 − F11 ]) =
1
(2)(2)
= .
(2 + 2 + 1)(2 + 2)
5
Therefore, due to Theorem 6.17,
E(Fy1 ) = E(F21 )E(F11 ) + E(F22 )E(1 − F11 )
µ ¶µ ¶ µ ¶µ ¶
1
1
1
1
1
=
+
= .
2
2
2
2
2
2
2
2
E(Fy1
) = E(F21
)E(F11
) + 2E(F21 )E(F22 )E(F11 [1 − F11 ])
2
)E([1 − F11 ]2 )
+E(F22
µ ¶µ ¶
µ ¶µ ¶µ ¶ µ ¶µ ¶
1
3
1
1
1
1
3
3
.
=
+2
+
=
3
10
2
2
5
3
10
10
So we have
V (Fy1 ) =
6.6.2
2
E(Fy1
)
3
−
− [E(Fy1 )] =
10
2
µ ¶2
1
= .05.
2
The Variance and Equivalent Sample Size
We introduce this subsection with an example.
Example 6.31 Consider the augmented Bayesian network in Figure 6.29 (b).
It is equivalent to the one in Figure 6.29 (a). Therefore, intuitively we would
2
) would be the same when computed using either network. Since
expect E(Fy1
clearly in that network Fy1 = F21 , we have due to Theorem 6.16 that
µ
¶µ
¶
2+1
2
3
2
2
) = E(F21
)=
E(Fy1
= ,
2+2+1
2+2
10
which is the same as the value obtained using Figure 6.29 (a). It is left as an
exercise to show the expected values are also equal, which means the variances
are equal.
2
) is not the same for the
It is also left as an exercise to show that E(Fy1
networks in Figures 6.29 (a) and (b) if we specify all beta(f ; 1, 1) density functions in both networks. Is there a theorem concerning equivalent sample sizes
and variances corresponding to Theorem 7.10? That is, is a variance the same
when computed using two equivalent augmented Bayesian networks? Although
we have no proof of this, we conjecture it is true. Before formally stating this
conjecture, we investigate more examples which substantiate it. We will only
6.6. VARIANCES IN COMPUTED RELATIVE FREQUENCIES
367
compare the expected values of squares of our random variables since, if they
are equal, the variances are equal. The reason is that the expected values of our
random variables are always equal to the corresponding probabilities. Before
giving a theorem to this effect, we motivate the theorem with two examples.
Example 6.32 Consider a two-node network such as the one in Figure 6.29
(a). Let Fy1 be a random variable which assigns P (y1|f) to each set of values f
of the variables in F. We have due to the law of total probability
Z
P (y1|f)ρ(f)df
E(Fy1 ) =
f
= P (y1).
Example 6.33 Consider again a two-node network such as the one in Figure
6.29 (a). Let Fx1|y1 be a random variable which assigns P (x1|y1, f) to each set
of values f of the variables in F. We have
Z
P (x1|y1, f)ρ(f|y1)df
E(Fx1|y1 |y1) =
f
= P (x1|y1).
The second equality is due to the law of total probability.
The method in the previous examples can be used to prove a theorem.
Theorem 6.18 Let a binomial augmented Bayesian network be given, and let
A and B be two disjoint subsets of V and a and b be values of the variables in
these sets. Let Fa|b be a random variable which assigns P (a|b, f) to each set of
values f of the variables in F. Then
E(Fa|b |b) = P (a|b).
Proof. We have
E(Fa|b |b) =
Z
P (a|b, f)ρ(f|b)df
f
= P (a|b).
The second equality is due to the law of total probability.
Since P (a|b) is the same for equivalent augmented Bayesian networks, the
preceding theorem implies the variances are the same whenever the expected
values of the squares of the random variables are the same. We now give more
examples comparing these expected values for equivalent augmented Bayesian
networks.
Example 6.34 Consider the augmented Bayesian networks in Figure 6.30 (a)
and (b).Clearly they are equivalent. For the network in (a) we have, due to
Theorem 6.16, that
3
3
=
E(F21 ) =
3+1
4
368
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
beta(f11; 2,2)
beta(f21; 1,1)
F11
F21
X
beta(f22; 1,1)
F22
Y
(a)
beta(f11; 1,1)
beta(f12; 1,1)
F11
beta(f21; 2,2)
F12
X
F21
Y
(b)
2
Figure 6.30: E(Fy1
) = 15/28 regardless of which network we use to compute it.
2
2
=
2+1
3
¶µ
¶
µ
5
4
4+1
2
=
)=
E(F11
4+3+1
4+3
14
µ
¶µ
¶
3+1
3
3
2
E(F21 ) =
=
3+1+1
3+1
5
µ
¶µ
¶
2+1
2
1
2
E(F22
)=
=
2+1+1
2+1
2
µ
¶µ
¶
3+1
3
3
E([1 − F11 ]2 ) =
=
4+3+1
4+3
14
E(F22 ) =
E(F11 [1 − F11 ]) =
3
(4)(3)
= .
(4 + 3 + 1)(4 + 3)
14
Therefore, due to Theorem 6.17,
2
2
2
) = E(F21
)E(F11
) + 2E(F21 )E(F22 )E(F11 [1 − F11 ])
E(Fy1
2
+E(F22
)E([1 − F11 ]2 )
µ ¶µ
¶
µ ¶µ ¶µ ¶ µ ¶µ ¶
3
5
3
2
3
1
3
15
=
+2
+
= .
5
14
4
3
14
2
14
28
6.6. VARIANCES IN COMPUTED RELATIVE FREQUENCIES
369
For the network in (b) we have, due to Theorem 6.16, that
µ
¶µ
¶
5+1
5
15
2
2
E(Fy1
.
) = E(F21
)=
=
5+2+1
5+2
28
The following theorem gives a formula for the expected value of the square
of the random variable. We will use the formula in several examples.
Theorem 6.19 Let a binomial augmented Bayesian network be given, and let
A and B be two disjoint subsets of V, and a and b be values of the variables in
these sets. Let Fa|b be a random variable which assigns P (a|b, f) to each set of
values f of the variables in F. Then
Z
1
[P (b|a, f)]2 [P (a|f)]2
2
ρ(f)df.
E(Fa|b
|b) =
P (b) f
P (b|f)
Proof. We have
2
E(Fa|b
|b) =
=
=
Z
[P (a|b, f)]2 ρ(f|b)df
f
Z
[P (b|a, f)]2 [P (a|f)]2 P (b|f)ρ(f)
df
[P (b|f)]2
P (b)
f
Z
1
[P (b|a, f)]2 [P (a|f)]2
ρ(f)df.
P (b) f
P (b|f)
Example 6.35 Consider the network in Figure 6.29 (a). Let Fx1|y1 be a random variable which assigns P (x1|y1, f) to each set of values f of the variables
in F. Due to the preceding theorem, we have
Z
1
[P (y1|x1, f)]2 [P (x1|f)]2
2
ρ(f)df
E(Fx1|y1 |y1) =
P (y1) f
P (y1|f)
Z 1Z 1Z 1
2 2
1
f21
f11
=
1/2 0 0 0 f21 f11 + f22 (1 − f11 )
beta(f11 ; 2, 2)beta(f21 ; 1, 1)beta(f22 ; 1, 1)df11 df21 df22
1
.
=
3
The integration was performed using the mathematics package Maple.
Consider next the augmented Bayesian network in Figure 6.29 (b), which is
equivalent to the one in (a). Due to the fact that F11 is independent of Y and
Theorem 6.16, we have
µ
¶µ
¶
1+1
1
1
2
2
|y1) = E(F11
)=
E(Fx1|y1
= ,
1+1+1
1+1
3
which is the same as the value above.
370
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
Example 6.36 Consider the network in Figure 6.30 (a). Let Fx1|y1 and Ey1
be as in the preceding example. Due to Theorem 6.19, we have
Z
1
[P (y1|x1, f)]2 [P (x1|f)]2
ρ(f)df
P (x2[1]) f
P (y1|f)
Z 1Z 1Z 1
2 2
1
f21
f11
=
5/7 0 0 0 f21 f11 + f22 (1 − f11 )
beta(f11 ; 4, 3)beta(f21 ; 3, 1)beta(f22 ; 2, 1)df11 df21 df22
2
.
=
5
2
|y1) =
E(Fx1|y1
The integration was performed using Maple.
Consider next the augmented Bayesian network in Figure 6.30 (b), which is
equivalent sample to the one in (a). Due to the fact that F11 is independent of
Y and Theorem 6.16, we have
µ
¶µ
¶
3+1
3
2
2
2
= .
E(Fx1|y1 |y1) = E(F11 ) =
3+2+1
3+2
5
which is the same as the value above.
Example 6.37 Consider the augmented Bayesian network in Figure 6.31 (a).
Let Fy1|x1,z1 be a random variable which assigns P (y1|x1, z1, f) to each set of
values f of the variables in F. Due to Theorem 6.19, we have
2
|x1, z1)
E(Fy1|x1,z1
=
=
=
=
=
Z
1
[P (x1, z1|y1, f)]2 [P (y1|f)]2
ρ(f)df
P (x1, z1) f
P (x1, z1|f)
Z
1
[P (x1, z1|y1, f)]2 [P (y1|f)]2
ρ(f)df
P (x1, z1) f P (x1, z1|y1, f)P (y1|f) + P (x1, z1|y2, f)P (y2|f)
Z
1
[P (z1|y1f)]2 [P (y1|x1, f)]2 P (x1|f)
ρ(f)df
P (x1, z1) f P (z1|y1, f)P (y1|x1, f) + P (z1|y2, f)P (y2|x1, f)
Z 1Z 1Z 1Z 1
2 2
1
f31
f21 f11
beta(f11 ; 1, 1)beta(f21 ; 1, 1)
1/4 0 0 0 0 f31 f21 + f32 (1 − f21 )
beta(f31 ; 1, 1)beta(f32 ; 1, 1)df11 df21 df31 df32
.36.
The third equality is obtained by exploiting the fact that X and Z are independent conditional on Y , using Bayes’ Theorem, and doing some manipulations.
The integration was performed using Maple.
Consider next the augmented Bayesian network in Figure 6.31 (b), which is
equivalent to the one in (a). Due to Theorem 6.19, we have
6.6. VARIANCES IN COMPUTED RELATIVE FREQUENCIES
beta(f11; 2,2)
beta(f21; 1,1)
F11
beta(f22; 1,1)
F21
F22
beta(f31; 1,1)
beta(f32; 1,1)
F31
Y
X
371
F32
Z
(a)
beta(f21; 2,2)
F21
beta(f11; 1,1)
beta(f31; 1,1)
Y
F31
F11
X
Z
beta(f12; 1,1)
beta(f32; 1,1)
F12
F32
(b)
2
Figure 6.31: E(Fy1|x1,z1
|x1, z1) = .36 regardless of which network we use to
compute it.
372
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
2
E(Fy1|x1,z1
|x1, z1)
Z
1
[P (x1, z1|y1, f)]2 [P (y1|f)]2
ρ(f)df
P (x1, z1) f
P (x1, z1|f)
Z 1Z 1Z 1Z 1Z 1
2 2 2
1
f21
f31 f11
=
beta(f11 ; 2, 2)
1/4 0 0 0 0 0 f21 f31 f11 + f22 f32 (1 − f11 )
beta(f21 ; 1, 1)beta(f22 ; 1, 1)beta(f31 ; 1, 1)beta(f32 ; 1, 1)df11 df21 df22 df31 df32
= .36,
=
which is the same as the value above. The integration was performed using
Maple.
Due to the preceding examples, we offer the following conjecture:
Conjecture 6.1 Suppose we have two equivalent binomial augmented Bayesian
network (G1 , F1 , ρ1 ) and (G2 , F2 , ρ2 ). Let A and B be two disjoint subsets of V,
and a and b be values of the variables in these sets. Furthermore, let F1,a|b be a
random variable which assigns P (a|b, f1 ) to each set of values f1 of the variables
in F1 , and let F2,a|b be a random variable which assigns P (a|b, f2 ) to each set of
values f2 of the variables in F2 . Then
2
2
|b) = E(F2,a|b
|b).
E(F1,a|b
Perhaps a proof technique, similar to the one in [Heckerman et al, 1995],
could be used to prove this conjecture.
6.6.3
Computing Variances in Larger Networks
So far we have computed variances involving only nodes that are touching each
other. Next we show how to compute variances in larger networks. We start
with an example.
Example 6.38 Consider the augmented Bayesian network in Figure 6.31 (a).
2
2
2
) and E(Fy1
). Let’s compute E(Fz1
).
We already know how to compute E(Fx1
We have
P (z1|f) = P (z1|y1, f)P (y1|f) + P (z1|y2, f)P (y2|f)
Therefore
Fz1 = F31 Fy1 + F32 (1 − Fy1 ).
This expression is like Expression 6.18 except Fy1 replaces F11 . So once we
2
), E(Fy1 ), E([1 − Fy1 ]2 ), and E(Fy1 [1 − Fy1 ]2 ), we can apply
determine E(Fy1
2
). From Example 6.30, we have
the method in Example 6.30 to compute E(Fz1
2
)=
E(Fy1
3
10
1
E(Fy1 ) = .
2
Due to symmetry
E([1 − Fy1 ]2 ) =
3
.
10
6.6. VARIANCES IN COMPUTED RELATIVE FREQUENCIES
373
Finally, We have
E(Fy1 [1 − Fy1 ]2 ) = E([F21 F11 + F22 {1 − F11 }][1 − F21 F11 − F22 {1 − F11 }])
1
.
=
5
The previous answer is obtained by multiplying out the expression on the right
and applying Theorem 6.16.
Note that all the values obtained above are exactly the ones that would be the
case if Fy1 had the beta(f21 ; 2, 2) density function. Therefore, due to the result
in Example 6.30,
3
2
E(Fz1
)= ,
10
which would be the value obtained if Fz1 had the beta(f31 ; 2, 2) density function.
The result in the previous example is not surprising since, if we took the
equivalent DAG that had the arrows reversed and used the same equivalent
sample size, Fz1 would have the beta(f21 ; 2, 2) density function. So the result is
consistent with Conjecture 6.1.
The previous example gives us insight as to how we can compute all prior
variances in a linked list. Starting from the root, at each node X, we compute
2
), E(Fx1 ), E([1 − Fx1 ]2 ), and E(Fx1 [1 − Fx1 ]2 ) from information obtained
E(Fx1
from the node above it and from the distributions of the auxiliary parent nodes
of X. Neapolitan and Kenevan [1990, 1991] extend this method to singlyconnected networks, and give a message-passing algorithm for computing all
prior variances in such networks. The algorithm is similar to the one discussed
in Section 3.2.2. They also show how to use Pearl’s ([Pearl, 1988]) method of
clustering to handle networks that are not singly-connected networks.
The problem of computing variances conditional on instantiated variables is
more difficult. Clearly, if an ancestor of a node is instantiated, the messagepassing algorithm described above can be used to compute the conditional variance. However, if a descendent is instantiated, we must cope with the integral in
Theorem 6.19. Che et al [1993] discuss using numerical methods to approximate
this integral.
6.6.4
When Do Variances Become Large?
Intuitively, we would expect the variance, in the random variable, whose possible values are a given conditional probability, could increase with the number
of instantiated variables relevant to the computation of that conditional probability. The reason is that, as more variables are instantiated, the number of
cases in the equivalent sample that have those instantiated values decreases.
For example, in Figure 6.32, all 80 cases enter into the determination of P (y1).
However, only 5 cases have X, Z, W , and U instantiated for x1, z1, w1, and u1.
Therefore, only 5 cases enter into the determination of P (y1|x1, z1, w1, u1). The
following table shows variances and 95% probability intervals given the network
in Figure 6.32:
374
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
beta(f11; 40,40)
X
beta(f21; 40,40)
Z
beta(f31; 40,40)
W
beta(f41; 40,40)
U
Y
for 1 # j # 16
beta(f5j; 2.5,2.5)
Figure 6.32: An augmented Bayesian network used to illustrate loss in confidence as variables are instantiated. Only the density functions for the auxiliary
variables are shown.
Random Variable
Fy1
Fy1|x1,z1,w1,u1
Expected Value
.5
.5
Variance
.003
.042
95% Probability Interval
(.391, .609)
(.123, .877)
Even though the expected value of the relative frequency with which Y = y1
remains at .5 when we instantiate Y ’s four parents, our confidence that it is .5
practically disappears.
When probabilities are nearly 1, we do not have as severe of a problem
concerning loss of confidence. The following theorem shows why.
Theorem 6.20 Let F be a random variable whose values range between 0 and
1. Then
V (F ) ≤ E(F )[1 − E(F )].
Proof. We have
V (F ) = E(F 2 ) − [E(F )]2
≤ E(F ) − [E(F )]2
≤ E(F )[1 − E(F )]
The second first inequality is because 0 ≤ f ≤ 1 implies f ≤ f 2 .
Example 6.39 Suppose E(F ) = .999. Then due to the preceding theorem,
V (F ) ≤ .999(1 − .999) = .000999.
6.6. VARIANCES IN COMPUTED RELATIVE FREQUENCIES
beta(f11; 40,40)
X
beta(f21; 40,40)
beta(f31; 40,40)
Z
375
beta(f41; 40,40)
W
U
Y
for 1 # j # 16
beta(f5j; 4.5,.5)
Figure 6.33: An augmented network used to illustrate loss in confidence as
variables are instantiated. Only the density functions for the auxiliary variables
are shown.
Using the normal approximation, we obtain that a 95% probability interval is
contained in
(.937, 1.061).
Of course, F cannot exceed 1. This is only an approximation.
So regardless of how many variables are instantiated, if the conditional probability (expected value of the relative frequency) is .999, we can be confident the
actual relative frequency really is high. Intuitively, the reason is that a relative
frequency estimate of .999 could not be based on a small sample.
However, if the probability is high but not extreme, we can lose a good deal
of confidence when we instantiate variables. Given the network in Figure 6.33,
we have the following variances and probability intervals:
Random Variable
Fy1
Fy1|x1,z1,w1,u1
Expected Value
.9
.9
Variance
.001
.015
95% Probability Interval
(.836, .964)
(.639, 1)
Note that, when Y ’s four parents are instantiated, we are no longer confident
that the relative frequency with which Y = y1 is high, even though the expected
value of the relative frequency stays at .9.
When variables are instantiated from above, as in the previous illustrations,
it is no surprise that the confidence becomes low because the confidence in
specified relative frequencies was low to begin with. However, when variables
are instantiated from below this is not the case. Consider the network in Figure
376
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
2
6.31 (b). As shown in Example 6.37, E(Fy1|x1,z1
|x1, z1) = .36. Therefore,
2
|x1, z1) − [E(Fy1|x1,z1 |x1, z1)]2
V (Fy1|x1,z1 |x1, z1) = E(Fy1|x1,z1
= .36 − (.5)2 = .11.
Yet the specified relative frequencies (i.e. F1i and F3j ), in which we have least
confidence, have the beta(f ; 1, 1) density function. So we have
V (F1i ) = V (F3j ) = .333 − (.5)2 = .083.
So even though E(Fy1|x1,z1 |x1, z1) is the same as the expected values of all
relative frequencies specified in the network (namely .5), its variance is greater
than any of their variances. It seems then that determination of variances may
be quite important when variables are instantiated from below. In this case,
we cannot assume that our confidence in specified relative frequencies gives
us a bound on our confidence in inferred ones. On the other hand, Henrion
et al [1996] note that diagnosis using Bayesian networks is often insensitive to
imprecision in probabilities. One reason they site is that gold-standard posterior
probabilities are often near zero or one, and, as we noted above, in the case of
extreme probabilities, the variance is always small.
EXERCISES
Section 6.1
Exercise 6.1 For some two-outcome experiment, which you can repeat indefinitely (such as the tossing of a thumbtack), determine the number of occurrences
a and b, of each outcome, that you feel your prior experience is equivalent to
having seen. Then represent your belief concerning the relative frequency with
the beta(f ; a, b) density function. Finally determine the probability of the first
value occurring.
Exercise 6.2 Assume I feel my prior experience concerning the relative frequency of smokers in a particular bar is equivalent to having seen 14 smokers
and 6 non-smokers. So I represent my beliefs concerning the relative frequency
of smokers using the beta(f ; 14, 6) density function. Suppose I then decide to log
whether or not individuals smoke. Compute the probability of my getting these
data:
{1, 2, 2, 2, 2, 1, 2, 2, 2, 1},
where 1 means the individual smokes and 2 means the individual does not smoke.
EXERCISES
377
Exercise 6.3 Assuming the beliefs in Exercise 6.2, what is the probability the
first individual sampled smokes? If we obtain the data shown in that exercise,
what is the updated beta density function representing my updated belief concerning the relative frequency of smokers? What is the probability the next individual
sampled smokes?
Section 6.2
Exercise 6.4 Suppose I am about to watch Sam and Dave race 10 times, and
Sam looks substantially athletically inferior to Dave. So I give Sam a probability
of .1 of winning the first race. However, I feel that if Sam wins once, he should
usually win. So given that Sam wins the first race, I give him a .8 probability of
winning the next one. Using the technique shown in Example 6.10, determine
the beta density function representing my prior belief concerning the relative
frequency with which Sam will win. Determine my probability of Sam winning
all 10 races.
Suppose next that Sam wins the first two races. Determine the updated beta
density function representing my updated belief concerning the relative frequency
with which Sam will win. Determine my probability of him winning the next race
and of winning all 8 remaining races.
Section 6.3
Exercise 6.5 Assuming the prior beliefs concerning the relative frequency of
smokers shown in Exercise 6.2, determine a 95% for the E(F ) where F is
a random variable representing my belief concerning the relative frequency of
smokers. Do the determination exactly and using the normal approximation.
Exercise 6.6 Assuming the prior beliefs concerning the relative frequency of
smokers shown in Exercise 6.2 and the data shown in that example, determine
a 95% for the E(F ) where F is a random variable representing my updated belief
concerning the relative frequency of smokers. Do the determination exactly and
using the normal approximation.
Section 6.4
Exercise 6.7 Show the Bayesian network embedded in the augmented Bayesian
network in Figure 6.34. Assume the random variables corresponding to the
conditional relative frequencies are as follows:
378
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
beta(f11; 4,8)
beta(f22; 6,6)
F11
F21
X1
X2
beta(f31; 1,1)
beta(f33; 1,3)
F31
F33
X3
beta(f32; 1,1)
beta(f34; 3,.1)
F32
F34
Figure 6.34: An augmented Bayesian network
Parent Values
X1 = 1, X2 = 1
X1 = 1, X2 = 2
X1 = 2, X2 = 1
X1 = 2, X2 = 2
RandomVarable
F31
F32
F33
F34
Exercise 6.8 Suppose we have a binomial Bayesian network sample whose parameter is the augmented Bayesian network in Figure 6.34, and we have these
data d:
Case
1
2
3
4
5
6
7
8
9
10
X1
1
1
2
2
1
2
1
2
1
1
X2
2
1
1
2
2
2
2
1
2
1
X3
1
2
1
1
1
2
1
2
1
1
Compute P (d) and ρ(fij |d) for all i, j. Show the updated augmented Bayesian
EXERCISES
379
X1
X2
X3
X4
Figure 6.35: A DAG.
network and updated embedded Bayesian network. Determine P (X3 = 1) for
the 11th case.
Exercise 6.9 Does the augmented Bayesian network in Figure 6.34 have an
equivalent sample size? If so, what is it?
Exercise 6.10 Complete the proof of Theorem 6.13.
Exercise 6.11 Use Theorem 6.13 to develop an augmented Bayesian network
with equivalent sample sizes 1, 2, 4, and 10 for the DAG in 6.35.
Exercise 6.12 Given the Bayesian network in Figure 6.36 and N = 36, use
Theorem 6.14 to create an augmented Bayesian network.
Exercise 6.13 Consider the augmented Bayesian network in Figure 6.37. What
is its equivalent sample size? Determine all augmented Bayesian networks that
are equivalent to it. Show that P (d) is the same for every Bayesian network
sample whose parameter is one of these augmented Bayesian networks, given
the data d in Exercise 6.8.
Section 6.5
Exercise 6.14 In the text, we updated the augmented Bayesian network in Figure 6.28 (a) with the data d in Table 6.4 using two iterations of Algorithm 6.1
(Expectation-Maximization). Starting with the results in the text, perform the
next two iterations.
Section 6.6
380
CHAPTER 6. PARAMETER LEARNING: BINARY VARIABLES
P(X1 = 1) = 3/4
P(X2 = 1) = 1/3
X1
P(X3 = 1|X1 = 1,X2 = 1) = 1/9
P(X3 = 1|X1 = 1,X2 = 2) = 1/2
X2
X3
P(X3 = 1|X1 = 2,X2 = 1) = 1/3
P(X3 = 1|X1 = 2,X2 = 2) = 1/6
X4
P(X4 = 1|X3 = 1) = 2/3
P(X4 = 1|X3 = 2) = 5/8
Figure 6.36: A Bayesian network.
beta(f21; 3,7)
beta(f11; 10,4)
F11
beta(f22; 1,3)
F21
X1
F22
beta(f31; 2,2)
F31
X2
beta(f32; 4,6)
F32
X3
Figure 6.37: An augmented Bayesian network.
Exercise 6.15 Consider the augmented Bayesian network in Figure 6.37. Let
F be a random variable representing our belief concerning the relative frequency
with which X2 equals 1. Compute E(F 2 ) using that network and all networks
equivalent to it. Are your values the same?
Chapter 7
More Parameter Learning
Chapter 6 considered Bayesian networks in which the variables are all binary. In
Section 7.1 we extend the theory presented in the previous chapter to multinomial variables. We provide fewer examples and intuitive explanations than usual
and we leave proofs of theorems and lemmas as exercises. The reason is that
the theory is a straightforward generalization of the theory for binary variables.
The notation is merely more difficult. After that, Section 7.2 discusses learning
parameters in the case of continuous variables.
7.1
Multinomial Variables
First we present the method for learning a single parameter in the case of a
multinomial variable; second we further discuss the Dirichlet density function;
third we show how to compute probability intervals and regions; and fourth we
present the method for learning all the parameters in a Bayesian network in
the case of multinomial variables. After all this, we close by briefly noting the
methods presented in Chapter 6 for learning parameters in the case of missing
data items and for computing variances in computed relative frequencies can
readily be extended to the case of multinomial variables.
7.1.1
Learning a Single Parameter
After discussing subjective probability distributions of relative frequencies in the
case of multinomial variables, we generalize the method developed in Section
6.1.2 for estimating relative frequencies from data.
Probability Distributions of Relative Frequencies
We start with a definition.
Definition 7.1 The Dirichlet density function with parameters a1 , a2 , . . . ar ,
381
382
CHAPTER 7. MORE PARAMETER LEARNING
4
2
0.2
0
0.2
0.6
0.6
0.8
0.8
1
1
Figure 7.1: The Dir(f1 , f2 ; 2, 2, 2) density function.
N=
Pr
k=1
ak , where a1 , a2 , . . . ar are integers ≥ 1, is
Γ(N )
a
f1 1−1 f2a2 −1 · · · frar −1
ρ(f1 , f2 , . . . fr−1 ) = Q
r
Γ(ak )
k=1
0 ≤ fk ≤ 1,
r
X
fk = 1.
k=1
Random variables F1 , F2 , . . . Fr , that have this density function, are said to have
a Dirichlet distribution.
The Dirichlet density function is denoted Dir(f1 , f2 , . . . fr−1 ; a1 , a2 , . . . ar ).
Note that the value of Fr isPuniquely determined by the values of the first
r−1
r − 1 variables (i.e. fr = 1 − h=1 fh ). That is why ρ is only a function of
r − 1 variables. As shown in Section 6.2.3, the Dirichlet density function is a
generalization of the beta density function. Figures 7.1 and 7.2 show Dirichlet
density functions.
As discussed in Section 6.2.3, there are cogent arguments for using the Dirichlet distribution to model our beliefs concerning relative frequencies. Often we
say the probability assessor’s experience is equivalent to having seen the kth
value occur ak times in N trials.
We will need the following lemma concerning the Dirichlet density function:
Lemma 7.1
PIf F1 , F2 , . . . Fr have a Dirichlet distribution with parameters a1 , a2 ,
. . . ar , N = ak , then for 1 ≤ k ≤ r
E(Fk ) =
ak
.
N
Proof. The proof is left as an exercise.
Now suppose we have some r-outcome random process. Let X be a random
variable whose space {1, 2, . . . r} contains the outcomes of the experiment, and
7.1. MULTINOMIAL VARIABLES
383
8
6
4
2
0.2
0
0.2
0.6
0.6
0.8
0.8
1
1
Figure 7.2: The Dir(f1 , f2 ; 4, 2, 2) density function.
for 1 ≤ k ≤ r let Fk be a random variable whose space is the interval [0, 1].
The probability distribution of Fk represents our belief concerning the relative
frequency with which X = k. Assume our beliefs are such that
P (X = k|fk ) = fk .
That is, if we knew for a fact that the relative frequency of k was fk , our belief
concerning the occurrence of k in the first execution of the experiment would
be fk . Given this assumption, the theorem that follows obtains our subjective
probability for the first trial.
Theorem 7.1 Suppose X is a random variable with space {1, 2, . . . r}, and
F1 , F2 , . . . Fr are r random variables such that for all k,
P (X = k|fk ) = fk .
Then
P (X = k) = E(Fk ).
Proof. The proof is left as an exercise.
Corollary 7.1 If the conditions in Theorem 8.2 hold, and P
if F1 , F2 , . . . Fr have
a Dirichlet distribution with parameters a1 , a2 , . . . ar , N = ak , then
ak
.
P (X = k) =
N
Proof. The proof follows immediately from Theorem 8.2 and Lemma 7.1.
384
CHAPTER 7. MORE PARAMETER LEARNING
Example 7.1 Suppose I am going to repeatedly throw a strange lop-sided die
with three sides. The shape is so odd that I have no reason to prefer one side
over the other, but, due to my lack of experience with such a die, I do not feel
strongly the relative frequencies are the same. So I model my beliefs with the
Dir(f1 , f2 ; 2, 2, 2) density function. We have
Γ(6)
f 2−1 f22−1 (1 − f1 − f2 )2−1
Γ(2)Γ(2)Γ(2) 1
= 120f1 f2 (1 − f1 − f2 ).
Dir(f1 , f2 ; 2, 2, 2) =
This density function is shown in Figure 7.1. Due to the previous corollary, for
the first throw we have
P (Side = 1) =
1
2
=
2+2+2
3
P (Side = 2) =
1
2
=
2+2+2
3
P (Side = 3) =
1
2
= .
2+2+2
3
Example 7.2 Suppose I am going to sample individuals in the United States,
and determine the relative frequency with which they wear colored, white, and
black socks. I think that about the same number of individuals wear black as
wear white socks, and about twice as many individuals wear colored sock as do
wear either of the former. However, I do not feel strongly about this belief. So
I model my beliefs with the Dir(f1 , f2 ; 4, 2, 2) density function. We have
Γ(8)
f 4−1 f22−1 (1 − f1 − f2 )2−1
Γ(4)Γ(2)Γ(2) 1
= 840f13 f2 (1 − f1 − f2 ).
Dir(f1 , f2 ; 4, 2, 2) =
This density function is shown in Figure 7.2. Due to the previous corollary, for
the first individual sampled,
P (Socks = colored) =
1
4
=
4+2+2
2
P (Socks = white) =
1
2
=
4+2+2
4
P (Socks = black) =
1
2
= .
4+2+2
4
Example 7.3 Suppose I am going to repeatedly throw an ordinary six-sided die.
I am fairly confident all relative frequencies are the same. So I model my beliefs
with the Dir(f1 , f2 , f3 , f4 , f5 ; 50, 50, 50, 50, 50, 50) density function.
7.1. MULTINOMIAL VARIABLES
385
Learning a Relative Frequency
We start with a definition.
Definition 7.2 Suppose we have a sample of size M such that
1. each X (h) has space {1, 2, . . . r};
2. F = {F1 , F2 , . . . Fr }, and for 1 ≤ h ≤ M and 1 ≤ k ≤ r
P (X (h) = k|f1 , . . . fk , . . . fr ) = fk .
Then D is called a multinomial sample of size M with parameter F.
Example 7.4 Suppose we throw a strange lop-sided die with three sides 10
times. Let k be the outcome if a k comes up, and let X (h) ’s value be the outcome
of the hth throw. Furthermore, let the density function for the variables in F be
Dir(f1 , f2 ; 2, 2, 2). Then D = {X (1) , X 2 , . . . X (10) } is a multinomial sample of
size 10 whose parameter has a Dirichlet distribution.
Before developing theory that enables us to update our belief about the next
trial from a multinomial sample, we present two more lemmas concerning the
Dirichlet distribution.
Lemma 7.2 Suppose FP
1 , F2 , . . . Fr have a Dirichlet distribution with parameP
ak , s1 , s2 , . . . sr are r integers ≥ 0, and M =
sk .
ters a1 , a2 , . . . ar , N =
Then
à r
!
r
Y s
Y
Γ(N )
Γ(ak + sk )
k
.
E
Fk
=
Γ(N + M )
Γ(ak )
k=1
k=1
Proof. The proof is left as an exercise.
Lemma 7.3 Suppose P
F1 , F2 , . . . Fr have a Dirichlet distribution with parameters a1 , a2 , . . . ar , N = ak , and s1 , s2 , . . . sr are r integers ≥ 0. Then
Q
( rk=1 f sk ) ρ(f1 , f2 , . . . fr−1 )
Qr
= Dir(f1 , f2 , . . . fr−1 ; a1 + s1 , a2 + s2 , . . . ar + sr ).
E ( k=1 Fksk )
Proof. The proof is left as an exercise.
Theorem 7.2 Suppose
1. D is a multinomial sample of size M with parameter F;
2. we have a set of values (data)
d = {x(1) , x(2) , . . . x(M ) }
of the variables in D;
386
CHAPTER 7. MORE PARAMETER LEARNING
3. sk is the number of variables in d equal to k.
Then
P (d) = E
Ã
r
Y
Fksk
k=1
!
.
Proof. The proof is left as an exercise.
Corollary 7.2 If the conditions in Theorem 7.2 hold, and
P F1 , F2 , . . . Fr have a
Dirichlet distribution with parameters a1 , a2 , . . . ar , N = ak , then
P (d) =
r
Γ(N ) Y Γ(ak + sk )
.
Γ(N + M)
Γ(ak )
k=1
Proof. The proof follows immediately from Theorem 7.2 and Lemma 7.2.
Example 7.5 Suppose we have the multinomial sample in Example 7.4, and
d = {1, 1, 3, 1, 1, 2, , 3, 1, 1, 1}.
Then a1 = a2 = a3 = 2, N = 6, s1 = 7, s2 = 1, s3 = 2, M = 10, and due to the
preceding corollary,
P (d) =
Γ(6) Γ(2 + 7) Γ(2 + 1) Γ(2 + 2)
= 4. 44 × 10−5 .
Γ(6 + 10) Γ(2)
Γ(2)
Γ(2)
Theorem 7.3 If the conditions in Theorem 7.2 hold, then
¶
µ r
Q sk
fk ρ(f1 , f2 , . . . fr−1 )
µ r
¶
,
ρ(f1 , f2 , . . . fr−1 |d) = k=1
Q sk
E
Fk
k=1
where ρ(f1 , f2 , . . . fr−1 |d) denotes the conditional density function of F1 , F2 ,
. . . Fr given D = d.
Proof. The proof is left as an exercise.
Corollary 7.3 Suppose the conditions in Theorem 7.2 hold, and
PF1 , F2 , . . . Fr
have a Dirichlet distribution with parameters a1 , a2 , . . . ar , N = ak . That is,
ρ(f1 , f2 , . . . fr−1 ) = Dir(f1 , f2 , . . . fr−1 ; a1 , a2 , . . . ar ).
Then
ρ(f1 , f2 , . . . fr−1 |d) = Dir(f1 , f2 , . . . fr−1 ; a1 + s1 , a2 + s2 , . . . ar + sr ).
Proof. The proof follows immediately from Theorem 7.3 and Lemma 7.3.
7.1. MULTINOMIAL VARIABLES
387
15
10
5
0.2
0
0.2
0.6
0.6
0.8
0.8
1
1
Figure 7.3: The Dir(f1 , f2 ; 9, 3, 4) density function.
The previous corollary shows that when we update a Dirichlet density function relative to a multinomial sample, we obtain another Dirichlet density function. For this reason, we say the set of all Dirichlet density functions is a
conjugate family of density functions for multinomial sampling.
Example 7.6 Suppose we have the multinomial sample in Example 7.4 and the
data in Example 7.5. Then a1 = a2 = a3 = 2, s1 = 7, s2 = 1, and s3 = 2. Due
to the preceding corollary,
ρ(f |d) = Dir(f1 , f2 ; 2 + 7, 2 + 1, 2 + 2) = Dir(f1 , f2 ; 9, 3, 4).
Figure 7.1 shows the original density function and Figure 7.3 shows the updated
density function.
Theorem 7.4 Suppose the conditions in Theorem 7.2 hold. If we create a
multinomial sample of size M + 1 by adding variable X (M +1) to D, then for
all k
P (X (M +1) = k|d) = E(Fk |d).
Proof. The proof is left as an exercise.
Corollary 7.4 If the conditions in Theorem 7.2 hold, and
PF1 , F2 , . . . Fr have a
Dirichlet distribution with parameters a1 , a2 , . . . ar , N = ak , then for all k
P (X (M +1) = k|d) =
ak + sk
.
N +M
388
CHAPTER 7. MORE PARAMETER LEARNING
Proof. The proof follows immediately from Theorem 7.4, Corollary 7.3, and
Lemma 7.1.
Example 7.7 Suppose we have the multinomial sample in Example 7.4 and the
data in Example 7.5. Then a1 = a2 = a3 = 2, N = 6, s1 = 7, s2 = 1, s3 = 2,
M = 10, and due to the preceding corollary,
P (X (M +1) = 1|d) =
2+7
= . 5625
6 + 10
P (X (M +1) = 2|d) =
2+1
= . 1875
6 + 10
P (X (M +1) = 3|d) =
7.1.2
2+2
= . 25.
6 + 10
More on the Dirichlet Density Function
After discussing the use of non-integral values in the Dirichlet density function,
we provide guidelines for assessing the values.
Non-integral Values of ak
So far we have only shown examples where ak is an integer ≥ 1 for each k. Figure
7.4 shows the Dir(f1 , f2 ; .2, .2, .2) density function. As that figure illustrates,
as all ak approach 0, we become increasingly certain the relative frequency of
one of the values is 1.
Assessing the Values of ak
Next we give some guidelines for choosing the size of ak in the Dirichlet distribution, when we are accessing our beliefs concerning a relative frequency.
• a1 = a2 = · · · = ar = 1: These values mean we consider all combinations
of relative frequencies that sum to 1 equally probable. We would use these
values when we feel we have no knowledge at all concerning the value of
the relative frequency. We might also use these values to try to achieve
objectivity in the sense that we impose none of our beliefs concerning
the relative frequency on the learning algorithm. We only impose the
fact that we know at most r things can happen. An example might be
learning the probability of low, medium, and high blood pressure from
data, which we want to communicate to the scientific community. The
scientific community would not be interested in our prior belief, but in
what the data had to say. Note we might not actually believe a priori
that all relative frequencies that sum to 1 are equally probable, but our
goal is not to impose this belief on the learning algorithm. Essentially the
posterior probability represents the belief of an agent that has no prior
beliefs concerning the relative frequencies.
7.1. MULTINOMIAL VARIABLES
389
3
2
1
0.2
0.6
0.8
1
0
0.2
0.6
0.8
1
Figure 7.4: The Dir(f1 , f2 ; .2, .2, .2) density function.
• a1 = a2 = · · · = ar > 1: These values mean we feel it more probable that
the relative frequency of the kth value is around ak /N . The larger the
values of ak , the more we believe this. We would use such values when
we want to impose our beliefs concerning the relative frequency on the
learning algorithm. For example, if we were going to toss a ordinary die,
we might take a1 = a2 = · · · = a6 = 50.
• a1 = a2 = · · · = ar < 1: These values mean we feel relative frequencies,
that result in not many different things happening, are more probable. We
would use such values when we want to impose our beliefs concerning the
relative frequencies on the system. For example, suppose we know there
are 1, 000, 000 different species, and we are about to land on an uncharted
island. We might feel it probable that not very many of the species are
present. So if we considered the relative frequencies with which we encountered different species, we would not consider relative frequencies, which
resulted in a lot of different species, probable. Therefore, we might take
ak = 1/1000, 000 for all k.
7.1.3
Computing Probability Intervals and Regions
We generalize the method for computing a probability interval developed in
Section 6.3.
Suppose F1 , F2 , . . . Fr , have a Dirichlet distribution. That is, their density
390
CHAPTER 7. MORE PARAMETER LEARNING
function is given by
Γ(N )
a
ρ(f1 , f2 , . . . fr−1 ) = Q
f1 1−1 f2a2 −1 · · · frar −1
r
Γ(ak )
0 ≤ fk ≤ 1,
k=1
r
X
fk = 1,
k=1
P
where a1 , a2 , . . . ar are integers ≥ 1, and N = rk=1 ak . By integrating over the
remaining variables, we obtain that the marginal density function of Fk is given
by
Γ(N )
a
f k−1 (1 − fk )bk −1 ,
ρ(fk ) =
Γ(ak )Γ(bk ) k
where
bk = N − ak .
It is left an exercise to show this. So
ρ(fk ) = beta(fk ; ak , bk ),
which means we can use all the techniques in Section 6.3 to compute a probability interval for Fk .
Example 7.8 Consider the Dir(f1 , f2 ; 4, 2, 2) density function, which was used
to model our beliefs concerning the color socks that people wear. We have N =
4 + 2 + 2 = 8. Therefore, due to Lemma 7.1, we have
E(F1 ) =
4
a1
= = .5.
N
8
Furthermore,
b1 = 8 − 4 = 4.
A 95% probability interval for F1 can therefore be found by solving the following
equation for c:
Z .5+c
Γ(8)
f14−1 (1 − f1 )4−1 df1 = .95.
Γ(4)Γ(4)
.5−c
Using the mathematics package Maple, we obtain the solution c = . 316. So our
95% probability interval for F1 is
(.5 − .316, .5 + .316) = (.184, .816).
Similarly, a 95% probability interval for F2 can be found by solving the following
equation for c:
Z .25+c
Γ(8)
f 2−1 (1 − f2 )6−1 df2 = .95.
Γ(2)Γ(6) 2
0
Using the mathematics package Maple, we obtain the solution c = . 271. So our
95% probability interval for F2 is
(0, .25 + .271) = (0, .521).
Clearly, this is also our 95% probability interval for F3 .
7.1. MULTINOMIAL VARIABLES
391
20
10
0.2
0
0.2
0.6
0.6
0.8
0.8
1
1
Figure 7.5: Given the Dir(f1 , f2 ; 8, 8, 16) density function, this is a 95% probability square for F1 and F2 .
We can also obtain probability regions for two or more random variables.
The following example illustrates this.
Example 7.9 Suppose we have the Dir(f1 , f2 ; 8, 8, 16) density function. We
can obtain a 95% probability square for F1 and F2 by solving the following equation for c:
Z
.25+c
.25−c
Z
.25+c
.25−c
Γ(32)
f 8−1 f28−1 (1 − f1 − f2 )16−1 df1 df2 = .95.
Γ(16)Γ(8)Γ(8) 1
Using the mathematics package Maple, the solution is c = . 167. Since .25 −
.167 = .0 83, and .25 + .167 = . 417, the following are corners of a square,
centered at (.25, .25), that contains 95% of the probability mass:
(.167, 167) (.417, .167)
(.167, .417)
(.417, .417).
Figure 7.5 shows this probability square.
When finding a region as in the previous example, we must be careful not
to cross the borders of where the density function is defined. For example, in
the case of three values, we must not only be careful to not cross each axis, but
also to not cross the line f1 + f2 = 1.
392
CHAPTER 7. MORE PARAMETER LEARNING
7.1.4
Learning Parameters in a Bayesian Network
Next we extend the theory for learning a single parameter to learning all the
parameters in a Bayesian network.
Multinomial Augmented Bayesian Networks
We start by generalizing Definition 6.9 of a binomial augmented Bayesian network.
Definition 7.3 A multinomial augmented Bayesian network (G, F, ρ) is
an augmented Bayesian network as follows:
1. For every i, Xi has space {1, 2, . . . ri }.
2. For every i there is an ordering [pai1 , pai2 , . . . paiqi ] of all instantiations
of the parents PAi in V of Xi , where qi is the number of different instantiations of these parents. Furthermore, for every i,
Fi = Fi1 ∪ Fi2 ∪ · · · Fiqi ,
where
Fij = {Fij1 , Fij2,... Fijri },
each Fij is a root, has no edge to any variable except Xi , and has density
function
ρij (fij ) = ρ(fij1 , fij2 , . . . fij(ri −1) )
0 ≤ fijk ≤ 1,
ri
X
fijk = 1.
k=1
3. For every i, j and k, and all values fi1 , . . . fij , . . . fiqi of Fi1 , . . . Fij , . . . Fiqi ,
P (Xi = k|paij , fi1 , . . . fij , . . . fiqi ) = fijk .
Since the Fij s are all root in a Bayesian network, they are mutually independent. So besides the Global Parameter Independence of the sets Fi , we have
Local Parameter Independence of their subsets Fij . That is, for 1 ≤ i ≤ n
ρ(fi1 , fi2 , . . . fiqi ) = ρ(fi1 )ρ(fi2 ) · · · ρ(fiqi ).
Global and local independence together imply
ρ(f11 , f12 , . . . fnqn ) = ρ(f11 )ρ(f12 ) · · · ρ(fnqn ).
(7.1)
Note that again to avoid clutter we did not subscript the density functions.
Note that a binomial augmented Bayesian network is a multinomial augmented Bayesian network in which ri = 2 for all i. Figure 7.6 shows a multinomial augmented Bayesian network that is not a binomial one. A multinomial augmented Bayesian network is a generalization of a binomial augmented
Bayesian network, and has all the same properties. To that effect, we have the
following theorem, which generalizes the corresponding theorem for binomial
augmented Bayesian networks.
7.1. MULTINOMIAL VARIABLES
Dir(f111,f112; 4,8,10)
F11
393
Dir(f211,f212,f213; 1,1,1,1)
Dir(f221,f222,f223; 2,4,1,1)
F21
F22
X1
X2
Dir(f231,f232,f233; 1,3,4,2)
F23
(a)
X1
X2
P(X1 = 1) = 2/11
P(X2 = 1|X1 = 1) = 1/4
P(X2 = 1|X1 = 3) = 1/10
P(X1 = 2) = 4/11
P(X2 = 2|X1 = 1) = 1/4
P(X1 = 3) = 5/11
P(X2 = 3|X1 = 1) = 1/4
P(X2 = 4|X1 = 1) = 1/4
P(X2 = 2|X1 = 3) = 3/10
P(X2 = 3|X1 = 3) = 2/5
P(X2 = 4|X1 = 3) = 1/5
P(X2 = 1|X1 = 2) = 1/4
P(X2 = 2|X1 = 2) = 1/2
P(X2 = 3|X1 = 2) = 1/8
P(X2 = 4|X1 = 2) = 1/8
(b)
Figure 7.6: A multinomial augmented Bayesian network is in (a), and its embedded Bayesian network is in (b).
Theorem 7.5 Let a multinomial augmented Bayesian network (G, F, ρ) be given.
Then for every i and j, the ijth conditional distribution in (G, P ) is given by
P (Xi = k|paij ) = E(Fijk ).
Proof. The proof is left as an exercise.
Corollary 7.5 Let a multinomial augmented Bayesian network be given. If
the variables in eachPFij have a Dirichlet distribution with parameters aij1 ,
aij2 , . . . aijri , Nij = k aijk , then for each i and each j, the ijth conditional
distribution in the embedded network (G, P ) is given by
P (Xi = k|paij ) =
aijk
.
Nij
Proof. The proof follows directly from Theorem 7.5 and Lemma 7.1.
394
CHAPTER 7. MORE PARAMETER LEARNING
Learning Using a Multinomial Augmented Bayesian Network
Next we give generalizations of the definitions and theorems in Section 6.4.3 to
multinomial augmented Bayesian networks. You are referred to that section for
discussions and examples, as we provide none here.
Definition 7.4 Suppose we have a Bayesian network sample of size M such
that
1. for every i each Xi(h) has space {1, 2, . . . ri };
2. its augmented Bayesian network (G, F, ρ) is multinomial.
Then D is called a multinomial Bayesian network sample of size M
with parameter (G, F).
Theorem 7.6 Suppose
1. D is a multinomial Bayesian network sample of size M with parameter
(G, F);
2. we have a set of values (data) of the X(h) s as follows:





(11
(2)
(M)
x1
x1
x1





. 
. 
.
x(2) = 
···
x(M ) = 
x(1) = 
 .. 
 .. 
 ..
(1)
(2)
(M)
xn
xn
xn




d = {x(1) , x(2) , . . . x(M ) };
3. Mij is number of x(h) s in which Xi ’s parents are in their jth instantiation,
and of these Mij cases, sijk is the number in which xi is equal to k.
Then
P (d) =
qi
n Y
Y
i=1 j=1
E
Ã
ri
Y
k=1
sijk
Fijk
!
.
Proof. The proof is left as an exercise.
Corollary 7.6 Suppose we have the conditions in Theorem 7.6 and the variables inPeach Fij have a Dirichlet distribution with parameters aij1 , aij2 , . . . aijri ,
Nij = k aijk . Then
P (d) =
qi
n Y
Y
ri
Y
Γ(Nij )
Γ(aijk + sijk )
.
Γ(N
+
M
)
Γ(aijk )
ij
ij
i=1 j=1
k=1
Proof. The proof follows immediately from Theorem 7.6 and Lemma 7.2.
7.1. MULTINOMIAL VARIABLES
395
Theorem 7.7 (Posterior Local Parameter Independence) Suppose we
have the conditions in Theorem 7.6. Then the Fij s are mutually independent
conditional on D. That is,
ρ(f11 , f12 , . . . fnqn |d) =
qi
n Y
Y
i=1 j=1
ρ(fij |d).
Furthermore,
ρ(fij |d) = ρ(fij1, fij2, . . . fij(ri −1) |d)
µr
¶
Qi sijk
fijk ρ(fij1 , fij2 , . . . fij(r1 −1) )
k=1
µr
¶
.
=
Qi sijk
E
Fijk
k=1
Proof. The proof is left as an exercise.
Corollary 7.7 Suppose we have the conditions in Theorem 7.6 and the variables inPeach Fij have a Dirichlet distribution with parameters aij1 , aij2 , . . . aijri ,
Nij = aijk . That is, for each i and each j
ρ(fij1 , fij2 , . . . fij(ri −1) ) = Dir(fij1 , fij2 , . . . fij(ri −1) ; aij1 , aij2 , . . . aijri ).
Then
ρ(fij1 , fij2 , . . . fij(ri −1) |d)
= Dir(fij1 , fij2 , . . . fij(ri −1) ; aij1 + sij1 , aij2 + sij2 , . . . aijri + sijri ).
Proof. The proof follows immediately from Theorem 7.7 and Lemma 7.3.
Using an Equivalent Sample Size
The results in Section 6.4.4 concerning equivalent sample sizes also hold for
multinomial augmented Bayesian networks. We just state the corresponding
results here.
Prior Equivalent Sample Size We start with a definition.
Definition 7.5 Suppose we have a multinomial augmented Bayesian network
in which the density functions are Dir(fij1 , fij2 , . . . fij(ri −1) ; aij1 , aij2 , . . . aijri )
for all i and j. If there is a number N such that for all i and j
Nij =
ri
X
k=1
aijk = P (paij ) × N,
then the network is said to have equivalent sample size N .
396
CHAPTER 7. MORE PARAMETER LEARNING
If a multinomial augmented Bayesian network has n nodes and equivalent
sample size N , we have for 1 ≤ i ≤ n,
qi
X
j=1
qi
qi
X
X
£
¤
P (paij ) × N = N ×
Nij =
P (paij ) = N.
j=1
j=1
It is unlikely we would arrive at a network with an equivalent sample size
simply by making up values of aijk . The next two theorems give common way
for constructing one.
Theorem 7.8 Suppose we specify G, F, and N and assign for all i, j, and k
aijk =
N
.
ri qi
Then the resultant augmented Bayesian network has equivalent sample size N ,
and the probability distribution in the resultant embedded Bayesian network is
uniform.
Proof. The proof is left as an exercise.
Theorem 7.9 Suppose we specify G, F, N, a Bayesian network (G, P ), and
assign for all i and j
aijk = P (Xi = k|paij ) × P (paij ) × N.
Then the resultant multinomial augmented Bayesian network has equivalent
sample size N . Furthermore, it embeds the originally specified Bayesian network.
Proof. The proof is left as an exercise.
Definition 7.6 Multinomial augmented Bayesian networks (G1 , F(G1 ) , ρ|G1 ) and
(G2 , F(G2 ) , ρ|G2 ) are called equivalent if they satisfy the following:
1. G1 and G2 are Markov equivalent.
2. The probability distributions in their embedded Bayesian networks are the
same.
3. The specified density functions in both are Dirichlet.
4. They have the same equivalent sample size.
Lemma 7.4 (Likelihood Equivalence) Suppose we have two equivalent multinomial augmented Bayesian networks (G1 , F(G1 ) , ρ|G1 ) and (G2 , F(G2 ) , ρ|G2 ). Let
D be a set of random vectors as specified in Definition 7.4. Then for every set
d of values of the vectors in D,
P (d|G1 ) = P (d|G2 )
where P (d|G1 ) and P (d|G2 ) are the probabilities of d when D is considered
multinomial Bayesian network samples with parameters (G1 , F(G1 ) ) and (G2 , F(G2 ) )
respectively.
Proof. The proof can be found in [Heckerman et al, 1995].
7.1. MULTINOMIAL VARIABLES
397
Theorem 7.10 Suppose we have two equivalent multinomial augmented Bayesian
networks (G1 , F(G1 ) , ρ|G1 ) and (G2 , F(G2 ) , ρ|G2 ). Let D be a set of random vectors as specified in Definition 7.4. Then given any set d of values of the vectors
in D, the updated embedded Bayesian network relative to D and the data d, obtained by considering D a binomial Bayesian network sample with parameter
(G1 , F(G1 ) ), contains the same probability distribution as the one obtained by
considering D a binomial Bayesian network sample with parameter (G2 , F(G2 ) ).
Proof. The proof is exactly like that of Theorem 7.10.
Corollary 7.8 Suppose we have two equivalent multinomial augmented Bayesian
networks (G1 , F(G1 ) , ρ|G1 ) and (G2 , F(G2 ) , ρ|G2 ). Then given any set d of values of the variables in D, the updated embedded Bayesian network relative to D
and the data d, obtained by considering D a binomial Bayesian network sample
with parameter (G1 , F(G1 ) ), is equivalent to the one obtained by considering D a
binomial Bayesian network sample with parameter (G2 , F(G2 ) ).
Proof. The proof follows easily from the preceding theorem.
Expressing Prior Indifference with a Prior Equivalent Sample Size
Recall in Section 6.4.4 we suggested that perhaps the best way to express prior
indifference, when every variable has two values, is simply to specify an equivalent sample size of two, and, for each node, distribute the sample evenly among
all specified values. If every variable has r values, this same argument suggests
we should use an equivalent sample size of r. However, in general, the variables
do not have the same number of values. So what size should we use in the
general case? It seems the most reasonable choice is to find the variable(s) with
the greatest number of values maxr, and use maxr as the equivalent sample
size. This may seem a bit strange because, for example, if X is a root with
two values and maxr = 16, then we specify a beta(f ; 8, 8) density function at
X. It seems that we have become highly confident P (X = 1) is equal to .5 just
because we included X in a network with other variables. The following provides a reasonable intuition justifying this. Suppose Y is a variable with maxr
values. In order to ‘know’ Y has maxr values, it is arguable that minimally
our prior experience must be equivalent to having seen each of them occur once.
Therefore, our prior sample size must be at least equal to maxr. Since X is in
that prior sample, there are also maxr observations of values of X.
Some Theoretical Results In Section 6.4.4, we argued for an equivalent
sample size on intuitive grounds. Furthermore, we proved we get the kind of
results we want when we use one. However, is there an axiomatic justification for assuming one? Heckerman et al [1995] discuss learning the conditional
independencies (and thereby a DAG pattern) among the variables from data,
a subject we discuss in Chapters 8-11. However, their results are relevant to
our considerations here. Namely, they show that if we make certain reasonable
assumptions, then we must use an equivalent sample size. Their assumptions
are the following: 1) we represent the belief that conditional independencies
are present using a multinomial augmented Bayesian network whose DAG G
398
CHAPTER 7. MORE PARAMETER LEARNING
entails all and only those conditional independencies; 2) the hypothesis that no
conditional independencies are present has positive probability; 3) when Xi has
the same set of parents in V in two multinomial augmented Bayesian networks,
the density functions of Fij for all j are the same in those networks (called Parameter Modularity); 4) if two multinomial augmented Bayesian networks
(G1 , F(G1 ) , ρ|G1 ) and (G2 , F(G2 ) , ρ|G2 ) satisfy only the first two conditions in
Definition 7.6, then we have Likelihood Equivalence (as defined in Lemma
7.4); and 5) all density functions are everywhere positive (i.e. the range of each
function includes only numbers greater than zero.). Given these assumptions,
they prove the density functions must be Dirichlet, and there is some N such
that each network has equivalent sample size N .
7.1.5
Learning with Missing Data Items
Algorithm 6.1, which appears in Section 6.5, extends immediately to one for
learning in the case of multinomial Bayesian networks.
7.1.6
Variances in Computed Relative Frequencies
It is left as an exercise to consult the references mentioned in Section 6.6 in
order to extend the results in that section to the case of multinomial variables.
Here we only state a generalization of Theorem 6.16, which you will need to
compute variances.
Theorem 7.11 Suppose F1 , F2 , . . . Fr have the Dir(f1 , f2 , . . . fr−1 ; a1 , a2 , . . . ar )
density function. Then
am
E(Fm ) = P
ak
µ
¶
¶µ
am + 1
a
2
P
E(Fm
)= P
ak + 1
ak
al am
P .
E(Fl Fm ) = P
( ak + 1) ak
Proof. The proof is left as an exercise.
7.2
Continuous Variables
Recall in Chapter 4.1 we defined a Gaussian Bayesian network, and we developed an algorithm for doing inference in such a network. Here we present a
method for learning parameters in Gaussian Bayesian networks. First we discuss
learning parameters for a normally distributed variable. Then we discuss learning parameters for variables which have the multivariate normal distribution.
Finally, we apply the theory developed to learning parameters in a Gaussian
Bayesian network.
7.2. CONTINUOUS VARIABLES
399
A
kA(a) = N(a;:,1/v)
X
kX(x|a) = N(x;a,1/r)
Figure 7.7: The probability distribution of A represents our belief concerning
the mean of X.
7.2.1
Normally Distributed Variable
First we assume the mean of the variable is unknown and the variance is known.
Then we discuss the case where the mean is known and the variance is unknown.
Finally, we assume both are unknown.
Before proceeding to do all this, we have the following definition:
Definition 7.7 Suppose X has the normal density function N (x; µ, σ2 ). Then
the precision r of X is defined as follows:
r=
1
.
σ2
Henceforth we show the normal distribution using the precision. The reason
should become clear as we develop the theory.
The Case of Unknown Mean and Known Variance
Suppose X is normally distributed with unknown mean and known precision r.
We represent our belief concerning the unknown mean with a random variable
A (for average), which is normally distributed with mean µ and precision v.
Note that the probability distribution of X is a relative frequency distribution
which models a phenomenon in nature, while the probability distribution of
A is our subjective probability concerning the value of X’s mean. Owing to
the discussion in Section 6.1.1, we represent this situation with the Bayesian
network in Figure 7.7.
The following theorem gives the prior density function of X:
Theorem 7.12 Suppose X and A are random variables such that
the density function of A is
ρA (a) = N (a; µ, 1/v),
400
CHAPTER 7. MORE PARAMETER LEARNING
kA(a) = N(a;:,1/v)
A
X(1)
X(2)
kX(1)(x(1)|a) = N(x(1);a,1/r)
X(M)
kX(2)(x(2)|a) = N(x(2);a,1/r)
kX(M)(x(M)|a) = N(x(M);a,1/r)
Figure 7.8: A Bayesian network representing a normal sample of size M with
parameter {A, r}.
and the conditional density function of X given A = a is
ρX (x|a) = N(x; a, 1/r).
Then the prior density function of X is
µ
¶
1 1
ρX (x) = N x; µ, +
.
r
v
Proof.
ρX (x) =
Z
ρX (x|a)ρA (a)da
a
=
Z
N (x; a, 1/r)N (a; µ, 1/v)da
a
=
Z
a
= N
N (a; x, 1/r)N (a; µ, 1/v)da
µ
x; µ,
1 1
+
r v
¶
The 3rd equality is due to Equality 4.2 and the 4th is due to Equality
4.5.
We see that X has a normal distribution with the same mean as A, but X
has greater variability than A owing to the uncertainty in X conditional on A’s
value.
Suppose now that we perform M trials of a random process whose outcome
is normally distributed with unknown mean and known precision r, we let X (h)
7.2. CONTINUOUS VARIABLES
401
be a random variable whose value is the outcome of the hth trial, and we let
A be a random variable representing our belief concerning the mean. Again
assume A is normally distributed with mean µ and precision v. As discussed in
Section 6.1.2, we assume that if we knew the value a of A for certain, then we
would feel the X (h) s are mutually independent, and our probability distribution
for each trial would have mean a. That is we have a sample defined as follows:
Definition 7.8 Suppose we have a sample of size M such that
1. each X (h) has space the reals;
2. F = {A, r},
ρA (a) = N (a; µ, 1/v)
and for 1 ≤ h ≤ M,
ρX (h) (x(h) |a) = N (x(h) ; a, 1/r).
Then D is called a normal sample of size M with parameter {A, r}.
We represent this sample with the Bayesian network in Figure 7.8
Theorem 7.13 Suppose
1. D is a normal sample of size M with parameter {A, r} where r > 0, and
A has precision v > 0;
2. d = {x(1) , x(2) , . . . x(M ) } is a set of values (data) of the variables in D,
and
PM (h)
x
.
x = h=1
M
Then the posterior density function of A is
ρA (a|d) = N (a; µ∗ , 1/v∗ )
where
µ∗ =
vµ + Mrx
v + Mr
Proof. It is left as exercises to show
"
and
v∗ = v + M r.
#
M
r X (h)
ρD (d|a) w exp −
(x − a)2 ,
2
(7.2)
(7.3)
h=1
where exp(y) denotes ey and w means ‘proportionate to’, and that
M
X
h=1
(x(h) − a)2 = M (a − x)2 +
M
X
(x(h) − x)2 .
h=1
(7.4)
402
CHAPTER 7. MORE PARAMETER LEARNING
Since the far right term in Equality 7.4 does not contain a, we may rewrite
Relation 7.3 as follows:
·
¸
Mr
(a − x)2 .
ρD (d|a) w exp −
(7.5)
2
The prior density function of A satisfies the following:
i
h v
ρA (a) w exp − (a − µ)2 .
2
We have
ρA (a|d) w ρD (d|a)ρA (a)
·
¸
i
h v
Mr
' exp − (a − µ)2 exp −
(a − x)2 .
2
2
(7.6)
(7.7)
The first proportionality in Relation 7.7 is due to Bayes’ Theorem, and the
second is due to Relations 7.5 and 7.6. It is left as an exercise to show
v(a − µ)2 + M r(a − x)2 = (v + Mr)(a − µ∗ )2 +
vM r(x − µ)2
.
v + Mr
(7.8)
Since the final term in Equality 7.8 does not contain a, it can also be included
in the proportionality factor. So we can rewrite Relation 7.7 as
·
¸
v + Mr
∗ 2
(a − µ )
ρA (a|d) w exp −
2
·
¸
(a − µ∗ )2
w exp −
2 (1/v∗ )
∗
(7.9)
w N (a; µ , 1/v ∗ ).
Since ρA (a|d) and N(a; µ∗ , 1/v ∗ ) are both density functions, their integrals over
the real line must both equal 1. Therefore, owing to Relation 7.9, they must be
the same function.
Example 7.10 Suppose D is a normal sample of size M with parameter {A, 1}.
That is, we represent our prior belief concerning A using a value of r = 1. Then,
owing to the previous theorem, our posterior density function of A is given by
ρA (a|d) = N (a; µ∗ , 1/v ∗ ),
where
µ∗ =
vµ + M x
v+M
and
v∗ = v + M.
(7.10)
From the previous example, we see that when r = 1, if we consider the
parameter µ the mean of the hypothetical sample on which we base our prior
belief concerning the value of A and v its size, then µ∗ would be the mean of
the hypothetical sample combined with the actual sample, and v ∗ would be the
size of this combined sample. Therefore, we can attach the following meaning
to the parameters µ and v:
7.2. CONTINUOUS VARIABLES
403
The parameter µ is the mean of the hypothetical sample upon which we base
our prior belief concerning the value of A.
The parameter v is the size of the hypothetical sample on which we base our
prior belief concerning the value of A.
Example 7.11 Suppose r = 1, v = 4 and µ = 10. Then we can consider our
prior belief as being equivalent to having seen a sample of size 4 in which the
mean was 10. Suppose next we sample 3 items, and we obtain x(1) = 4, x(2) = 5,
x(4) = 6. Then M = 3, x = (4 + 5 + 6)/3 = 5, and therefore owing to Equality
7.10
(4 × 10) + (3 × 5)
vµ + M x
=
= 7.86
µ∗ =
v+M
4+3
and
v ∗ = 4 + 3 = 7.
Example 7.12 Note that v must be greater than 0 because if v were 0, the
variance of A would be infinite. We can take the limit as v → 0 (i.e. as the
variance approaches infinity) of the expressions for µ∗ and v∗ (Equality 7.2) to
model the situation in which we have complete prior ignorance as to the value
of A. Taking these limits we have
µ∗ = lim
v→0
vµ + M rx
=x
v + Mr
and
v ∗ = lim (v + M r) = M r.
v→0
Alternatively, we could obtain the same result using an improper prior
density function. It is called improper because it is not really a density function in that its integral is not finite. We obtain an improper density function
by taking the limit of the prior density function as v → 0. The prior density
function appears in Equality 7.6. Taking the limit of that density function, we
have
³
h v
i´
ρA (a) w lim exp − (a − µ)2 = 1.
v→0
2
Note that this is only a formal result because we are ignoring the constant of
proportionality which is approaching 0. This improper density function imparts
a uniform distribution over the whole real line.
Using this improper prior density function and proceeding from Relation 7.7,
we obtain that
ρA (a|d) = N (a; x, 1/Mr),
which is the same as the result obtained by taking the limit as v → 0 of the
expressions for µ∗ and v ∗ .
404
CHAPTER 7. MORE PARAMETER LEARNING
Assuming the improper prior density function, the posterior variance of A
is σ2 /M , where σ2 = 1/r is the known variance of each X i conditional on A.
A posterior perc % probability interval for A is therefore given by
¶
µ
σ
σ
x − zperc √ , x + zperc √
,
M
M
where zperc is the z-score obtained from the Standard Normal Table (See Section
6.3 for a discussion of probability intervals and the z-score.). If you are familiar with ‘classical’ statistics, you should notice that this probability interval is
identical to the perc % confidence interval for the mean of a normal distribution
with know variance σ2 .
Note that if we take v = 0, then the prior density function of X (obtained
in Theorem 7.12) is also improper.
Next we give a theorem for the density function of X (M +1) , the M + 1st trial
of the experiment.
Theorem 7.14 Suppose we have the assumptions in Theorem 7.13.
X (M +1) has the posterior density function
µ
¶
1
(M +1)
(M +1)
∗ 1
|d) = N x
;µ , + ∗ ,
ρX (M+1) (x
r v
Then
where the values of µ∗ and v ∗ are those obtained in Theorem 7.13.
Proof. The proof is left as an exercise.
The Case of Known Mean and Unknown Variance
Next we discuss the case where the mean is known and the variance is unknown.
First we need to review the gamma distribution.
The Gamma Distribution We start with the following definition:
Definition 7.9 The gamma density function with parameters α and β,
where α > 0 and β > 0, is
ρ(x) =
β α α−1 −βx
x
e
Γ(α)
x > 0,
and is denoted gamma(x; α, β).
A random variables X that has this density function is said to have a gamma
distribution.
If the random variable X has the gamma density function, then
E(X) =
α
β
and
V (X) =
α
.
β2
7.2. CONTINUOUS VARIABLES
405
0.12
0.1
0.08
0.06
0.04
0.02
0
2
4
6
x8
10
12
14
Figure 7.9: The chi-square density function with 6 degrees of freedom.
Definition 7.10 The gamma density function is called the chi-square (χ2 )
density function with k degrees of freedom, where k is a positive integer,
when
1
k
and
β= ,
α=
2
2
2
and is denoted chi-square(x; k) and χk (x).
A random variables X that has this density function is said to have a chisquare (χ2 ) distribution with k degrees of freedom.
Example 7.13 We have
chi-square(x; 6) =
=
=
β α α−1 −βx
x
e
Γ(α)
(1/2)6/2 6/2−1 −(1/2)x
x
e
Γ (6/2)
1 2 −x/2
x e
.
16
Figure 7.9 shows this density function.
The following theorem is a well-known result concerning the gamma density
function.
Theorem 7.15 Suppose X1 , X2 , . . . Xk are k independent random variables,
each with the N (x; 0, σ2 ) density function. Then the random variable
V = X12 + X22 + · · · Xk2
has the gamma(v, k/2, 1/2σ2 ) density function.
Proof. The proof is left as an exercise.
406
CHAPTER 7. MORE PARAMETER LEARNING
Corollary 7.9 Suppose X1 , X2 , . . . Xk are k independent random variables, each
with the standard normal density function. Then the random variable
V = X12 + X22 + · · · Xk2
has the chi-square(v; k) density function.
Proof. The proof follows immediately from the preceding theorem.
Learning With Known Mean and Unknown Variance Now we can discuss learning in the case of known mean and unknown variance. Our goal is
to proceed quickly to the case where both are unknown. So the only result
we present here is a theorem which obtains the posterior distribution of the
variance. We obtain this result because we refer to it in the next subsection.
Suppose we perform M trials of a random process whose outcome is normally
distributed with known mean a and unknown variance, and we let X (h) be a
random variable whose value is the outcome of the hth trial. We represent our
belief concerning the unknown precision with a random variable R, which has
the gamma(r; α/2, β/2) density function. Similar to before, we assume that if
we knew the value r of R for certain, then we would feel the X (h) s are mutually
independent, and our probability distribution for each trial would have precision
r. That is we have a sample defined as follows:
Definition 7.11 Suppose we have a sample of size M such that
1. each X (h) has space the reals;
2. F = {a, R},
ρR (r) = gamma (r; α/2, β/2) ,
and for 1 ≤ h ≤ M
ρX (h) (x(h) |r) = N (x(h) ; a, 1/r).
Then D is called a normal sample of size M with parameter {a, R}.
The following theorem obtains the updated distribution of R given this sample.
Theorem 7.16 Suppose
1. D is a normal sample of size M with parameter {a, R};
2. d = {x(1) , x(2) , . . . x(M ) } is a set of values of the variables in D, and
s=
M ³
´2
X
x(h) − a .
h=1
7.2. CONTINUOUS VARIABLES
407
Then the posterior density function of R is
ρR (r|d) = gamma(r, α∗ /2, β ∗ /2)
where
β∗ = β + s
and
α∗ = α + M.
(7.11)
Proof. The proof is left as an exercise.
Let’s investigate the parameters α and β. Suppose x̂(1) , x̂(2) , . . . x̂α are the
values in the hypothetical sample upon which we base our prior belief concerning
the value of R, and β is the value of s for that sample. That is,
β=
α
X
¡ i
¢2
x̂ − a .
i=1
∗
Then clearly β would be the value of s for the combined sample. Therefore,
we see that we can attach the following meaning to the parameters α and β:
The parameter β is the value of s in the hypothetical sample upon which we
base our prior belief concerning the value of R.
The parameter α is the size of the hypothetical sample upon which we base
our prior belief concerning the value of R.
Example 7.14 We can take the limit as β → 0 and α → 0 of the expressions
for α∗ and β ∗ to model the situation in which we have complete prior ignorance
as to the value of R. Taking these limits we have
β ∗ = lim (β + s) = s
β→0
α∗ = lim (α + M) = M.
α→0
Alternatively, we could obtain the same result by taking the limit of the prior
density function gamma (r; α/2, β/2) as β → 0 and α → 0. Taking the limit of
that density function, we have
¡
¢ 1
ρR (r) w lim lim rα−1 e−βr = .
α→0 β→0
r
This improper density function assigns a large probability to large variances
(small values of r), and small probability to small variances.
Using this improper prior density function, it is left as an exercise to show
ρR (r|d) = gamma(r; M/2, s/2).
which is the same as the result obtained by taking the limit as β → 0 and α → 0
of the expressions for β ∗ and α∗ .
408
CHAPTER 7. MORE PARAMETER LEARNING
0.35
0.3
0.25
0.2
0.15
0.1
0.05
-4
-2
0
2 x
4
Figure 7.10: The t density function with 3 degrees of freedom.
The Case of Unknown Mean and Unknown Variance
We now assume both the mean and the variance are unknown. First we need
to review the t distribution.
The t Distribution We start with the following definition:
Definition 7.12 The t density function with α degrees of freedom, where
α > 0, is
ρ(x) =
µ
1
απ
¶1/2
¢µ
¡
¶− (α+1)
2
Γ α+1
x2
2
¡α¢ 1 +
α
Γ 2
− ∞ < x < ∞,
(7.12)
and is denoted t(x; α).
A random variables X that has this density function is said to have a t
distribution with α degrees of freedom.
If the random variable X has the t(x; α) density function and if α > 2, then
E(X) = 0
and
V (X) =
α
.
α−2
Figure 7.10 shows the t density function with 3 degrees of freedom. Note its
similarity to the standard normal density function. Indeed, it is left as an
exercise to show the standard normal density function is equal to the limit as α
approaches infinity of the t distribution with α degrees of freedom.
The family of t distributions can be enlarged so that it includes every density
function which can be obtained from a density function of the form shown in
Equality 7.12 through an arbitrary translation and change of scale. We do that
next.
7.2. CONTINUOUS VARIABLES
409
Definition 7.13 A random variables X has a t distribution with α degrees of
freedom, where α > 0, location parameter µ, where −∞ < µ < ∞, and precision
τ , where τ > 0, if the random variable τ 1/2 (X − µ) has a t distribution with α
degrees of freedom. The density function of X is
i− (α+1)
³ τ ´ 1 Γ ¡ α+1 ¢ h
τ
2
2
¡ α2 ¢ 1 + (x − µ)2
ρ(x) =
απ
α
Γ 2
− ∞ < x < ∞, (7.13)
and is denoted t(x; α, µ, τ ).
If the random variable X has the t(x; α, µ, τ ) density function and if α > 2,
then
α
E(X) = µ
and
V (X) =
τ −1 .
α−2
Note that the precision in the t distribution does not exactly equal the inverse
of the variance as it does in the normal distribution. The N (x; µ, 1/τ) is equal
to the limit as α approaches infinity of the t(x; α, µ, τ ) density function (See
[DeGroot, 1970].).
Learning With Unknown Mean and Unknown Variance Now we can
discuss learning when both the mean and the variance are unknown. Suppose
X is normally distributed with unknown mean and unknown precision. We
again represent our belief concerning the unknown mean and unknown precision with the random variables A and R respectively. We assume R has the
gamma (r; α/2, β/2) density function and A has the N (a; µ, 1/vr) conditional
density function. The following theorem gives the prior density function of X.
Theorem 7.17 Suppose X , A, and R are random variables such that
the density function of R is
ρR (r) = gamma (r; α/2, β/2) ,
the conditional density function of A given R = r is
ρA (a|r) = N (a; µ, 1/vr) ,
and the conditional density function of X given A = a and R = r is
ρX (x|a, r) = N(x; a, 1/r).
Then the prior density function of X is
µ
ρX (x) = t x; α, µ,
vα
(v + 1) β
¶
.
(7.14)
410
CHAPTER 7. MORE PARAMETER LEARNING
Proof. We have
ρX (x) =
=
Z Z
a r
Z Z
a
ρX (x|a, r)ρA (a|r)ρR (r)drda
N (x; a, 1/r)N (a; µ, 1/vr) gamma (r; α/2, β/2) drda.
r
It is left as an exercise to perform this integration and obtain Equality 7.14.
Suppose now that we perform M trials of a random process whose outcome
is normally distributed with unknown mean and unknown variance, we let X (h)
be a random variable whose values are the outcomes of the hth trial, and we
represent our belief concerning each trial as in Theorem 7.17. As before, we
assume that if we knew the values a and r of A and R for certain, then we
would feel the X (h) s are mutually independent, and our probability distribution
for each trial would have mean a and precision r. That is we have a sample
defined as follows:
Definition 7.14 Suppose we have a sample of size M such that
1. each X (h) has space the reals;
2. F = {A, R},
ρR (r) = gamma (r; α/2, β/2) ,
ρA (a|r) = N (a; µ, 1/vr) ,
and for 1 ≤ h ≤ M
ρX (h) (x(h) |a, r) = N(x(h) ; a, 1/r).
Then D is called a normal sample of size M with parameter {A, R}.
Theorem 7.18 Suppose
1. D is a normal sample of size M with parameter {A, R};
2. d = {x(1) , x(2) , . . . x(M ) } is a set of values of the variables in D, and
x=
PM
h=1
M
x(h)
and
M ³
´2
X
x(h) − x .
s=
h=1
Then the posterior density function of R is
ρR (r|d) = gamma(r, α∗ /2, β ∗ /2)
where
β∗ = β + s +
vM (x − µ)2
v+M
and
α∗ = α + M,
(7.15)
7.2. CONTINUOUS VARIABLES
411
and the posterior conditional density function of A given R = r is
ρA (a|r, d) = N (a; µ∗ , 1/v∗ r)
where
vµ + Mx
and
v∗ = v + M.
v+M
Proof. It is left as exercises to show
"
#
M
r X (h)
n/2
2
exp −
(x − a) ,
ρD (d|a, r) w r
2
µ∗ =
(7.16)
h=1
and that
ρA,R (a, r) = ρA (a|r)ρR (r)
w r
(7.17)
1/2 −(vr/2)(a−µ)2 α/2−1 −βr/2
e
r
e
.
Owing to Bayes Theorem, ρA,R (a, r|d) is proportional to the product of the right
sides of Relations 7.16 and 7.17. Using equalities similar to Equalities 7.4 and
7.8, it is left as an exercise to perform steps similar to those in Theorem 7.13
to obtain
½
· ∗
¸¾ n
o
∗
∗
v r
1/2
∗ 2
rα /2−1 e−β /2 .
(a − µ )
ρA,R (a, r|d) w r exp
2
This completes the proof.
Based on a discussion similar to that following Example 7.10, we can attach
the following meaning to the parameters µ and v:
The parameter µ is the mean of the hypothetical sample upon which we base
our prior belief concerning the value of A.
The parameter v is the size of the hypothetical sample upon which we base
our prior belief concerning the value of A.
Let’s investigate the parameters α and β. Suppose x̂(1) , x̂(2) , . . . x̂v are the
values in the hypothetical sample upon which we base our prior belief concerning
the value of A, µ is the mean of that sample, and β is the value of s in that
sample. That is,
v
X
¡ i
¢2
x̂ − µ .
(7.18)
β=
i=1
Based on these assumptions, it is left as an exercise to use the left equality in
Equalities 7.15 to show
β∗ =
v
M ³
´2
X
¡ i
¢2 X
x(h) − µ∗ ,
x̂ − µ∗ +
i=1
(7.19)
h=1
which would be the value of s for the combined sample. Therefore, we see that
we can attach the following meaning to the parameter β:
412
CHAPTER 7. MORE PARAMETER LEARNING
The parameter β is the value of s in the hypothetical sample upon which we
base our prior belief concerning the value of A.
There does not seem to be a clear-cut meaning we can attach to the parameter α. Based on the right equality in Equality 7.15 and the result obtained
in Section 7.2.1 for the case where only the mean is known, we may want to
say it is about equal to the size of the hypothetical sample upon which we base
our prior belief concerning the value of R. However, Equality 7.18 indicates we
should make v that size. These results suggest we should make α about equal
to v. Indeed, the next example shows that is reasonable to make α equal to
v − 1.
Example 7.15 Suppose we want to express complete prior ignorance as to the
values of A and R by using the product of the improper density function when
we knew only the value of the mean and the improper density function when we
knew only the value of the variance. That is, we take the product of the uniform
density function over the whole real line (which is simply 1) and the function
1/r. We then have
1
ρA,R (a, r) = .
r
Recall Relation 7.17 says
2
ρA,R (a, r) w r1/2 e−(vr/2)(a−µ) rα/2−1 e−βr/2 .
In order for the limit of this expression to be equal to 1/r, we would have to take
the limit as v → 0, β → 0, and α → −1. If we use these values to model prior
ignorance, we obtain
β∗ = s
µ∗ = x
and
and
α∗ = M − 1,
v∗ = M ,
the posterior density function of R is
ρR (r|d) = gamma(r; (M − 1) /2, s/2),
and the posterior conditional density function of A given R = r is
ρA (a|r, d) = N (a; x, 1/M r).
It is left as an exercise to show that this means sr = s/σ 2 is distributed
χ (M − 1). We therefore have
h
i
s
P χ2P1 (M − 1) < 2 < χ2P2 (M − 1) = P2 − P1 ,
σ
2
where χ2P i (M − 1) is the Pi fractional point of the χ2 (M − 1) distribution. A
few manipulations yields
·
¸
s
s
2
<
σ
<
P
= P2 − P1 .
χ2P2 (M − 1)
χ2P1 (M − 1)
7.2. CONTINUOUS VARIABLES
413
If you are familiar with ‘classical’ statistics, you should notice that this P2 − P1
probability interval for σ2 is identical to the P2 − P1 confidence interval for the
variance of a normal distribution with unknown mean and unknown variance.
Next we give a theorem for the density function of X (M +1) , the M + 1st trial
of the experiment:
Theorem 7.19 Suppose we have the assumptions in Theorem 7.18.
X (M +1) has the posterior density function
µ
¶
v ∗ α∗
(M +1)
(M +1)
∗
∗
ρX (M+1) (x
|d) = t x
;α ,µ , ∗
,
(v + 1) β ∗
Then
where the values of α∗ , β ∗ , µ∗ , and v ∗ are those obtained in Theorem 7.18.
Proof. The proof is left as an exercise.
7.2.2
Multivariate Normally Distributed Variables
After reviewing the multivariate normal, the Wishart, and the multivariate t
distributions, we obtain our result for learning parameters when variables have
the multivariate normal distribution. We prove few results in this subsection.
Most are generalizations of the results in the previous subsection.
The Multivariate Normal Distribution
Next we discuss the mulitivariate normal distribution, which is a generalization
of the normal distribution to more than one variable. In this context, we call
the normal distribution the univariate normal distribution and the distribution of two variables the bivariate normal distribution. We discuss this latter
distribution first in order to make it easier to understand the general case.
Bivariate Normal Distribution Defined
Definition 7.15 The bivariate normal density function with parameters
µ1 , σ 1 , µ2 , σ2 , and p, where −∞ < µi < ∞, σi > 0, and |p| < 1, is
ρ(x1 , x2 ) =
1
×
2πσ1 σ2 (1 − p2 )1/2
"µ
(
µ
¶
¶ #)
x1 − µ1 2
1
(x1 − µ1 ) (x2 − µ2 )
x2 − µ2 2
exp −
− 2p
+
2(1 − p2 )
σ1
σ1 σ2
σ2
−∞ < xi < ∞, and is denoted N (x1 , x2 ; µ1 , σ21 , µ2 , σ22 , p).
Random variables X1 and X2 that have this density function are said to have
the bivariate normal distribution.
414
CHAPTER 7. MORE PARAMETER LEARNING
0.15
0.1
0.05
0
-4
-4
-2
-2
0
0
2
2
4
4
Figure 7.11: The N (x1 , x2 ; 0, 1, 0, 1, 0) density function.
If the random variables X1 and X2 have the bivariate normal density function, then
E(X1 ) = µ1
and
V (X1 ) = σ21 ,
E(X2 ) = µ2
and
V (X2 ) = σ22 ,
and
p (X1 , X2 ) = p,
where p (X1 , X2 ) denotes the correlation coefficient of X1 and X2 .
Example 7.16 We have
2
2
N (x1 , x2 ; 0, 1 , 0, 1 , 0) =
=
·
¸
¢
1¡ 2
1
2
exp − x1 + x2
2π
2
2
x1
x22
1 −
1 −
√ e 2 √ e 2 ,
2π
2π
which is the product of two standard univariate normal density functions. This
density function, which appears in Figure 7.11, is called the bivariate standard
normal density function.
Example 7.17 We have
7.2. CONTINUOUS VARIABLES
415
0.006
0.004
0.002
0
0
-20
0
20
20
40
Figure 7.12: The N(x1 , x2 ; 1, 2, 20, 12, .5) density function.
N(x1 , x2 ; 1, 22 , 20, 122 , .5) =
1
×
2π(2)(12)(1 − .52 )1/2
(
"µ
µ
¶2
¶2 #)
x2 − 20
1
(x1 − 1) (x2 − 20)
x1 − 1
+
exp −
− 2(.5)
.
2(1 − .52 )
2
(2)(12)
12
Figure 7.12 shows this density function.
In Figures 7.11 and 7.12, note the familiar bell-shaped curve which is characteristic of the normal density function. The following two theorems show the
relationship between the bivariate normal and the normal density functions.
Theorem 7.20 If X1 and X2 have the N(x1 , x2 ; µ1 , σ21 , µ2 , σ22 , p) density function, then the marginal density function of X1 is
ρX1 (x1 ) = N (x1 , ; µ1 , σ21 ).
Proof. The proof is developed in the exercises.
Theorem 7.21 If X1 and X2 have the N(x1 , x2 ; µ1 , σ21 , µ2 , σ22 , p) density function, then the conditional densify function of X1 given X2 = x2 is
ρX1 (x1 |x2 ) = N (x1 ; µX1 |x2 , σ2X1 |x2 ),
where
µX1 |x2 = µ1 + p
µ
σ1
σ2
¶
(x2 − µ2 )
416
CHAPTER 7. MORE PARAMETER LEARNING
and
σ2X1 |x2 = (1 − p2 )σ21 .
Proof. The proof is left as an exercise.
More on Vectors and Matrices Recall we defined random vector and random matrix in Section 5.3.1. Before proceeding, we discuss random vectors
further. Similar to the discrete case, in the continuous case the joint density
function of X1 , . . . and Xn is represented using a random vector as follows:
ρX (x) ≡ ρX1 ,...Xn (x1 , . . . xn ).
We call


E(X1 )


..
E(X) ≡ 

.
E(Xn )
the mean vector of random vector X, and

V (X1 )
Cov(X1 , X2 )
 Cov(X2 , X1 )
V (X2 )

Cov(X) ≡ 
..
..

.
.
···
···
..
.
Cov(Xn , X1 ) Cov(Xn , X2 ) · · ·
Cov(X1 , Xn)
Cov(X2 , Xn)
..
.
V (Xn , Xn )





the covariance matrix of X. Note that the covariance matrix is symmetric.
We often denote a covariance matrix as follows:

 2
σ1 σ12 · · · σ1n
 σ21 σ22 · · · σ2n 


ψ= .
..
..  .
..
 ..
.
.
. 
σn1
σn2
···
σ2n
Recall that the transpose XT of column vector X is the row vector defined
as follows:
¢
¡
XT = X1 · · · Xn .
We have the following definitions:
Definition 7.16 A symmetric n × n matrix a is called positive definite if
xT ax > 0
for all n-dimensional vectors x 6= 0, where 0 is the vector with all 0 entries.
Definition 7.17 A symmetric n× n matrix a is called positive semidefinite
if
xT ax ≥ 0
for all n-dimensional vectors x.
7.2. CONTINUOUS VARIABLES
417
Recall a matrix a is called non-singular if there exists a matrix b such that
ab = I, where I is the identity matrix. Otherwise it is called singular??. We
have the following theorem:
Theorem 7.22 If a matrix is positive definite, then it is nonsingular; and if a
matrix is positive semidefinite but not positive definite, then it is singular.
Proof. The proof is left as an exercise.
Example 7.18 The matrix
µ
1 0
0 1
µ
1 1
1 1
is positive definite. You should show this.
Example 7.19 The matrix
¶
¶
is positive semidefinite but not positive definite. You should show this.
Multivariate Normal Distribution Defined We can now define the multivariate normal distribution.
Definition 7.18 Let


X1


X =  ... 
Xn
be a random vector. We say X has a multivariate normal distribution if
for every n-dimensional vector bT ,
bT X
either has a univariate normal distribution or is constant.
The previous definition does not give much insight into multivariate normal
distributions or even if one exists. The following theorems show they do indeed
exist.
Theorem 7.23 For every n-dimensional vector µ and n × n positive semidefinite symmetric matrix ψ, there exists a unique multivariate normal distribution
with mean vector µ and covariance matrix ψ.
Proof. The proof can be found in [Muirhead, 1982].
Owing to the previous theorem, we need only specify a mean vector µ and
a positive semidefinite symmetric covariance matrix ψ to uniquely obtain a
multivariate normal distribution. Theorem 7.22 implies that ψ is nonsingular
if and only if it is positive definite. Therefore, if ψ is positive definite, we say
the distribution is a nonsingular multivariate normal distribution, and
otherwise we say it is a singular multivariate normal distribution. The
next theorem gives us a density function for the nonsingular case.
418
CHAPTER 7. MORE PARAMETER LEARNING
Theorem 7.24 Suppose the n-dimensional random vector X has a nonsingular
multivariate normal distribution with mean vector µ and covariance matrix ψ.
Then X has the density function
·
¸
1 2
1
exp − ∆ (x) ,
ρ(x) =
2
(2π)n/2 (det ψ)1/2
where
∆2 (x) = (x − µ)T ψ−1 (x − µ).
This density function is denoted N(x; µ, ψ).
Proof. The proof can be found in [Flury, 1997].
The inverse matrix
T = ψ−1
is called the precision matrix of N (x; µ, ψ). If µ = 0 and ψ is the identity matrix, N (X;µ, ψ) is called the multivariate standard normal density
function.
Example 7.20 Suppose n = 2 and we have the multivariate standard normal
density function. That is,
µ ¶
0
µ=
0
and
ψ=
µ
1 0
0 1
Then
T = ψ−1 =
µ
1
0
¶
.
0
1
¶
,
∆2 (x) = (x − µ)T ψ−1 (x − µ)
¶
µ
¶µ
¡
¢ 1 0
x1
x
x
=
1
2
x2
0 1
= x21 + x22 ,
and
N (x; µ, ψ) =
=
=
=
·
¸
1 2
exp − ∆ (x)
2
(2π)n/2 (det ψ)1/2
·
¸
¡
¢
1 2
1
2
x
exp
−
+
x
2
2 1
(2π)2/2 (1)1/2
·
¸
¢
1¡
1
exp − x21 + x22
2π
2
N (x1 , x2 ; 0, 12 , 0, 12 , 0),
1
7.2. CONTINUOUS VARIABLES
419
which is the bivariate standard normal density function.
It is left as an exercise to show that in general if
¶
µ
µ1
µ=
µ2
and
ψ=
is positive definite, then
µ
σ 21
σ21
σ12
σ22
¶
N (x; µ, ψ) = N (x1 , x2 ; µ1 , σ21 , µ2 , σ22 , σ 12 / [σ1 σ2 ]).
Example 7.21 Suppose
µ=
µ
µ
1
1
and
ψ=
3
3
¶
1
1
¶
.
Since ψ is not positive definite, Theorem 7.24 does not apply. However, since
ψ is positive semidefinite, Theorem 7.23 says there is a unique multivariate
normal distribution with this mean vector and covariance matrix. Consider the
distribution of X1 and X2 determined by the following density function and
equality:
(x1 − 3)2
1 −
2
ρ(x1 ) = √ e
2π
X2 = X1 .
Clearly this distribution has the mean vector and covariance matrix above. Furthermore, it satisfies the condition in Definition 7.18. Therefore, it is the unique
multivariate normal distribution that has this mean vector and covariance matrix.
Note in the previous example that X has a singular multivariate normal distribution, but X1 has a nonsingular multivariate normal distribution. In general,
if X has a singular multivariate normal distribution, there is some linear relationship among the components X1 , . . . Xn of X, and therefore these n random
variables cannot have a joint n-dimensional density function. However, if some
of the components are deleted until there are no linear relationships among
the ones that remain, then the remaining components will have a nonsingular
multivariate normal distribution.
Generalizations of Theorems 7.20 and 7.21 exist. That is, if X has the
N (X; µ, ψ) density function, and
¶
µ
X1
,
X=
X2
420
CHAPTER 7. MORE PARAMETER LEARNING
then the marginal distribution of X1 and the conditional distribution of X1
given X2 = x2 are both multivariate normal. You are referred to [Flury, 1997]
for statements and proofs of these theorems.
The Wishart Distribution
We have the following definition:
Definition 7.19 Suppose X1 , X2 , . . . Xk are k independent n-dimensional random vectors, each having the multivariate normal distribution with n-dimensional
mean vector 0 and n × n covariance matrix ψ. Let V denote the random symmetric k × k matrix defined as follows:
V = X1 XT1 + X2 XT2 + · · · Xk XTk .
Then V is said to have a Wishart distribution with k degrees of freedom and
parametric matrix ψ.
Owing to Theorem 7.22, ψ is positive definite if and only if it is nonsingular. If k > n − 1 and ψ is positive definite, the Wishart distribution is called
nonsingular. In this case, the precision matrix T of the distribution is defined
as
T = ψ−1 .
The follow theorem obtains a density function in this case:
Theorem 7.25 Suppose n-dimensional random vector V has the nonsingular
Wishart distribution with k degrees of freedom and parametric matrix ψ. Then
V has the density function
·
¸
¢
1 ¡
ρ(v) = c (n, k) |ψ|−k/2 |v|(k−n−1)/2 exp − tr ψ−1 v ,
2
where tr is the trace function and
"
c (n, k) = 2
kn/2 n(n−1)/4
π
µ
¶#−1
n
Y
k+1−i
Γ
.
2
i=1
(7.20)
This density function is denoted W ishart(v; k, T).
Proof. The proof can be found in [DeGroot, 1970].
It is left as an exercise to show that if n = 1, then W ishart(v; k, 1/σ2 ) =
gamma(v; k/2, 1/2σ2 ). However, showing this is not really necessary because it
follows from Theorem 7.15 and the definition of the Wishart distribution.
7.2. CONTINUOUS VARIABLES
421
The Multivariate t Distribution
We have the following definition:
Definition 7.20 Suppose n-dimensional random vector Y has the N (Y; µ, ψ)
density function, T = ψ−1 , random variable Z has the chi−square(z; α) density
function, Y and Z are independent, and µ is an arbitrary n-dimensional vector.
Define the n-dimensional random vector X as follows: For i = 1, . . . n
Xi = Yi
µ ¶−1/2
Z
+ µi .
α
Then the distribution of X is called a multivariate t distribution with α
degrees of freedom, location vector µ, and precision matrix T.
The following theorem obtains the density function for the multivariate t
distribution.
Theorem 7.26 Suppose n-dimensional random vector X has the multivariate
t distribution with α degrees of freedom, location vector µ, and precision matrix
T. Then X has the following density function:
¸−(α+n)/2
·
1
T
,
ρ(x) = b (n, α) 1 + (x − µ) T(x − µ)
α
where
b (n, α) =
Γ
¡ α+n ¢
2
|T|1/2
Γ (α/2) (απ)n/2
(7.21)
.
This density function is denoted t(x; α, µ, T) .
Proof. The proof can be found in [DeGroot, 1970].
It is left as an exercise to show that in the case where n = 1 the density
function in Equality 7.21 is the univariate t density function which appears in
Equality 7.13.
If the random vector X has the t(x; α, µ, T) density function and if α > 2,
then
α
T−1 .
E(X) = µ
and
Cov(X) =
α−2
Note that the precision matrix in the t distribution is not the inverse of the
covariance matrix as it is in the normal distribution. The N (x; µ, T−1 ) is equal
to the limit as α approaches infinity of the t(x; α, µ, T) density function (See
[DeGroot, 1970].).
Learning With Unknown Mean Vector and Unknown Covariance Matrix
We discuss the case where both the mean vector and the covariance matrix
are unknown. Suppose X has a multivariate normal distribution with unknown
422
CHAPTER 7. MORE PARAMETER LEARNING
mean vector and unknown precision matrix. We represent our belief concerning the unknown mean vector and unknown precision matrix with the random vector A and the random matrix R respectively.
´ R has the
³ We assume
−1
conditional
W ishart(r; α, β) density function and A has the N a; µ, (vr)
density function. The following theorem gives the prior density function of X.
Theorem 7.27 Suppose X and A are n-dimensional random vectors, and R
is an n × n random matrix such that
the density function of R is
ρR (r) = W ishart(r; α, β)
where α > n − 1 and β is positive definite (i.e. the distribution is nonsingular).
the conditional density function of A given R = r is
´
³
ρA (a|r) = N a; µ, (vr)−1
where v > 0,
and the conditional density function of X given A = a and R = r is
ρX (x|a, r) = N (x; a, r−1 ).
Then the prior density function of X is
µ
¶
v(α − n + 1) −1
ρX (x) = t x; α − n + 1, µ,
β
.
(v + 1)
(7.22)
Proof. The proof can be found in [DeGroot, 1970].
Suppose now that we perform M trials of a random process whose outcome
has the multivariate normal distribution with unknown mean vector and unknown precision matrix, we let X(h) be a random vector whose values are the
outcomes of the hth trial, and we represent our belief concerning each trial as
in Theorem 7.27. As before, we assume that if we knew the values a and r of
A and R for certain, then we would feel the X (h) s are mutually independent,
and our probability distribution for each trial would have mean vector a and
precision matrix r. That is, we have a sample defined as follows:
Definition 7.21 Suppose we have a sample of size M as follows:
1. We have the n-dimensional random vectors




X1(1)
X1(2)
 . 
 . 


X(2) = 
X(1) = 
 .. 
 .. 
(1)
(2)
Xn
Xn
···
D = {X(1) , X(2) , . . . X(M ) }
(h)
such that for every i each Xi
has space the reals.
X(M )

X1(M)


..

=
.


(M)
Xn

7.2. CONTINUOUS VARIABLES
2. F = {A, R},
423
ρR (r) = W ishart(r; α, β),
where α > n − 1 and β is positive definite,
´
³
−1
ρA (a|r) = N a; µ, (vr)
where v > 0, and for 1 ≤ h ≤ M
ρX(h) (x(h) |a, r) = N(x(h) ; a, r−1 ).
Then D is called multivariate normal sample of size M with parameter
{A, R}.
The following theorem obtains the updated distributions of A and R given
this sample.
Theorem 7.28 Suppose
1. D is a multivariate normal sample of size M with parameter {A, R};
2. d = {x(1) , x(2) , . . . x(M) } is a set of values of the random vectors in D, and
x=
PM
h=1 x
M
(h)
and
M ³
´³
´T
X
x(h) − x x(h) − x .
s=
h=1
Then the posterior density function of R is
ρR (r|d) = W ishart(r; α∗ , β ∗ )
where
β∗ = β + s +
vM
(x − µ)(x − µ)T
v+M
α∗ = α + M,
and
(7.23)
and the posterior conditional density function of A given R = r is
−1
ρA (a|r, d) = N (a; µ∗ , (v ∗ r)
where
µ∗ =
vµ + M x
v+M
and
)
v∗ = v + M.
Proof. The proof can be found in [DeGroot, 1970].
As in the univariate case which is discussed in Section 7.2.1, we can attach
the following meaning to the parameters:
The parameter µ is the mean vector in the hypothetical sample upon which
we base our prior belief concerning the value of A.
424
CHAPTER 7. MORE PARAMETER LEARNING
The parameter v is the size of the hypothetical sample upon which we base
our prior belief concerning the value of A.
The parameter β is the value of s in the hypothetical sample upon which we
base our prior belief concerning the value of A.
It seems reasonable to make α equal to v − 1.
Similar to the univariate case, we can model prior ignorance by setting v = 0,
β = 0, and α = −1 in the expressions for β ∗ , α∗ , µ∗ , and v ∗ . However, we
must also assume M > n. See [DeGroot, 1970] for a complete discussion of this
matter. Doing so, we obtain
β∗ = s
and
α∗ = M − 1,
and
µ∗ = x
v ∗ = M.
and
Example 7.22 Suppose n = 3, we model prior ignorance by setting v = 0,
β = 0, and α = −1, and we obtain the following data:
Case
1
2
3
4
X1
1
5
2
8
X2
2
8
4
6
X3
6
2
1
3
Then M = 4 and
x(1)


1
= 2 
6
x(2)


5
= 8 
2
x(3)


2
= 4 
1
x(4)
So

      
1
5
2
8
 2 + 8 + 4 + 6 
6
2
1
3
x =
4
 
4
=  5 
3


8
=  6 .
3
7.2. CONTINUOUS VARIABLES
and
So
and
425



−3
1
¡
¢
s =  −3  −3 −3 3 +  3
3
−1



−2
¡
¢
+  −1  −2 −1 −2 + 
−2


30
18
−6
=  18
20 −10  .
−6 −10 14

30
β ∗ = s =  18
−6

18
−6
20 −10 
−10 14


4
µ∗ = x =  5 
3
and
and


¡
1 3 −1

4 ¡
1  4 1
0
0
¢
¢
α∗ = M − 1 = 3,
v ∗ = M = 4.
Next we give a theorem for the density function of X(M+1) , the M +1st trial
of the experiment.
Theorem 7.29 Suppose we have the assumptions in Theorem 7.28. Then
X(M +1) has the posterior density function
µ
¶
v ∗ (α∗ − n + 1) ∗ −1
(β
)
ρX(M+1) (x(M +1) |d) = t x(M+1) ; α∗ − n + 1, µ∗ ,
,
(v ∗ + 1)
where the values of α∗ , β∗ , µ∗ , and v∗ are those obtained in Theorem 7.28.
Proof. The proof is left as an exercise.
7.2.3
Gaussian Bayesian Networks
A Gaussian Bayesian network uniquely determines a nonsingular multivariate
normal distribution and vice versa. So to learn parameters for a Gaussian
Bayesian network we can apply the theory developed in the previous subsection.
First we show the transformation; then we develop the method for learning
parameters.
Transforming a Gaussian Bayesian Network to a Multivariate Normal
Distribution
Recall that in Section 4.1.3 a Gaussian Bayesian network was defined as follows.
If PAX is the set of all parents of X, then
X
x = wX +
bXZ z,
(7.24)
Z∈PAX
426
CHAPTER 7. MORE PARAMETER LEARNING
where WX has density function N (w; 0, σ2WX ), and WX is independent of each
Z. The variable WX represents the uncertainty in X’s value given values of X’s
parents. Recall further that σ2WX is the variance of X conditional on values of
its parents. For each root X, its unconditional density function N (x; µX , σ2X )
is specified.
We will show how to determine the multivariate normal distribution corresponding to a Gaussian Bayesian network; but first we develop a different
method for specifying a Gaussian Bayesian network. We will consider a variation
of the specification shown in Equality 7.24 in which each WX does not necessarily have zero mean. That is, each WX has density function N (w; E(WX ), σ2WX ).
Note that a network, in which each of these variables has zero mean, can be
obtained from a network specified in this manner by giving each node X an auxiliary parent Z, which has mean E(WX ), zero variance, and for which bXZ = 1.
If the variable WX in our new network is then given a normal density function
with zero mean and same variance as the corresponding variable in our original
network, the two networks will contain the same probability distribution.
Before we develop the new way we will specify Gaussian Bayesian networks,
recall that an ancestral ordering of the nodes in a directed graph is an ordering
of the nodes such that if Y is a descendent of Z, then Y follows Z in the ordering.
Now assume we have a Gaussian Bayesian network determined by specifications
as in Equality 7.24, but in which each WX does not necessary have zero mean.
Assume we have ordered the nodes in the network according to an ancestral
ordering. Then each node is a linear function of the values of all the nodes that
precede it in the ordering, where some of the coefficients may be 0. So we have
xi = wi + bi1 x1 + bi2 x2 + · · · bi,i−1 xi−1 ,
where Wi has density function N (wi ; E(Wi ), σ2i ), and bij = 0 if Xj is not a
parent of Xi . Then the conditional density function of Xi is
X
bij xj , σ2i ).
(7.25)
ρ(xi |pai ) = N (xi ; E(Wi ) +
Xj ∈PAi
Since
E(Xi ) = E(Wi ) +
X
bij E(Xj ),
(7.26)
Xj ∈PAi
we can specify the unconditional mean of each variable Xi instead of the unconditional mean of Wi . So our new way to specify a Gaussian Bayesian
network is to show for each Xi its unconditional mean µi ≡ E(Xi ) and its
conditional variance σ2i . Owing to Equality ??, we have then
X
E(Wi ) = µi −
bij µj .
Xj ∈PAi
Substituting this expression for E(Wi ) into Equality 7.25, we have that the
conditional density function of Xi is
X
bij (xj − µj ), σ2i ).
(7.27)
ρ(xi |pai ) = N (xi , µi +
Xj ∈PAi
7.2. CONTINUOUS VARIABLES
427
F12
:1
F22
:2
b21
X1
X2
Figure 7.13: A Gaussian Bayesian network.
Figures 7.13, 7.14, and 7.15 show examples of specifying Gaussian Bayesian
networks in this manner.
Next we show how we can generate the mean vector and the precision matrix
for the multivariate normal distribution determined by a Gaussian Bayesian
network. The method presented here is from [Shachter and Kenley, 1989]. Let
and
ti =
1
,
σ2i

bi1
..
.

bi = 

bi,i−1


The mean vector in the multivariate normal distribution corresponding to a
Gaussian Bayesian network is simply


µ1


µ =  ...  .
µn
The following algorithm creates the precision matrix.
T1 = (t1 ) ;
for (i =µ2; i <= n; i + +)
Ti−1 + ti bi bTi
Ti =
−ti bTi
T = Tn ;
// Determine the precision matrix T.
−ti bi
ti
¶
;
// The precision matrix is Tn .
Example 7.23 We apply the previous algorithm to the Gaussian Bayesian network in Figure 7.13. We have
T1 = (t1 ) = (1/σ 1 )
428
CHAPTER 7. MORE PARAMETER LEARNING
F12
:1
F22
:2
X1
X2
b31
b32
X3
F32
:3
Figure 7.14: A Gaussian Bayesian network.
T2
µ
¶
T1 + t2 b2 bT2 −t2 b2
−t2 bT2
t2


³ ´
³ ´
1
1
1
(b
(b
+
)
(b
)
−
)
2
2
2
21
21
21
σ
σ2

³σ2 ´
=  1
1
− σ12 (b21 )
σ22
2
Ã
!
2
b
1
+ σ212 − bσ212
σ21
2
2
=
.
1
− bσ212
σ2
=
2
2
So the multivariate normal distribution determined by this Gaussian Bayesian
network has mean vector
¶
µ
µ1
µ=
µ2
and precision matrix
T=
Ã
2
+ bσ212
2
− bσ212
1
σ21
2
− bσ212
1
σ22
2
!
.
It is left as an exercise to show the covariance matrix is
µ
¶
σ21
b21 σ21
−1
ψ=T =
.
b21 σ21 σ22 + b221 σ21
(7.28)
7.2. CONTINUOUS VARIABLES
429
Example 7.24 We apply the preceding algorithm to the Gaussian Bayesian
network in Figure 7.14. First note that b21 = 0. We then have
T1 = (t1 ) = (1/σ 1 )
=
T2
µ

= 

= 
=
T3
=
µ


= 



= 



= 

Ã
¶
T1 + t2 b2 bT2 −t2 b2
−t2 bT2
t2

³ ´
³ ´
1
1
+ σ 2 (b21 ) (b21 ) − σ12 (b21 )
σ 21
2

³2 ´
1
− σ12 (b21 )
2
σ
³ ´ 2 
³ 2´
1
1
1
+ σ 2 (0) (0) − σ2 (0)
σ 21
2
³ 2´

1
− σ12 (0)
2
σ2
2
!
1
0
σ21
0 σ12
2
¶
T2 + t3 b3 bT3 −t3 b3
−t3 bT3
t3
à 1
!

³ ´µ b ¶
³ ´µ b ¶¡
0
¢
2
σ1
31
31
b31 b32
− σ12
+ σ12

b32
b32
3
3
0 σ12

2

³ ´¡
¢
1
b31 b32
− σ12
2
σ3
! 3
à 1
¶
¶ 
µ
µ
´
´
³
³
2
0
b
b
b
b
σ21
31 32
31
31
+ σ12
− σ12

b32 b31
b232
b32
3
3
0 σ12

2

³ ´¡
¢
1
b31 b32
− σ12
2
σ3
3

b231
b31 b32
b31
1
+ σ2
− σ2
σ 21
σ23
3
3

b232
b32 b31
1
+ σ 2 − bσ322 
.
σ23
σ22
3
3
b31
b32
1
− σ2
− σ2
σ2
3
3
3
So the multivariate normal distribution determined by this Gaussian Bayesian
network has mean vector


µ1
µ =  µ2 
µ3
and precision matrix


T=

b2
1
+ σ312
σ 21
3
b32 b31
σ23
− bσ312
3
b31 b32
σ23
b2
1
+ σ322
σ22
3
− bσ322
3
− bσ312

3

− bσ322 
.
3
1
σ23
430
CHAPTER 7. MORE PARAMETER LEARNING
F12
:1
F22
:2
b21
X1
X2
b31
b32
X3
F32
:3
Figure 7.15: A complete Gaussian Bayesian network.
It is left as an exercise to show the covariance matrix is


σ21
0
b31 σ21
.
σ22
b32 σ22
ψ = T−1 =  0
2
2
2 2
b31 σ1 b32 σ2 b31 σ1 + b232 σ 22 + σ23
(7.29)
Example 7.25 Suppose we have the Gaussian Bayesian network in Figure
7.15. It is left as an exercise to show that the multivariate normal distribution determined by this network has mean vector


µ1
µ =  µ2  ,
µ3
precision matrix

1
σ21
+
b221
σ22
b231
σ23
b32 b31
σ 23
+

b21
T=
 − σ 22 +
− bσ312
3
and covariance matrix
− bσ212 +
2
1
σ22
b31 b32
σ23
b232
σ23
+
− bσ322
3
− bσ312

3

− bσ322 
,
3
1
σ23
7.2. CONTINUOUS VARIABLES
431
ψ = T−1 =






σ21
b21 σ 21
b21 σ 21
σ22 + b221 σ21
(b31 + b32 b21 ) σ 21
b32 σ 22
¢
¡
2
+ b32 b21 + b31 b21 σ21

(b31 + b32 b21 ) σ21

b32 σ22 + ¢

¡
2

b32 b21 + b31 b21 σ21

2
2
2
2 2

σ
+
b
σ
+
b
σ
32 2¢
¡ 23 2 31 1
+ b21 b32 + 2b21 b31 b32 σ21
(7.30)
Learning Parameters in a Gaussian Bayesian network
First we define a Gaussian augmented Bayesian network.
Definition 7.22 A Gaussian augmented Bayesian network (G, F, ρ) is
an augmented Bayesian network as follows:
1. For every i, Xi is a continuous random variable.
2. For every i,
Fi = {Bi , Mi , Σ2i },
where BTi = (Bi1 , . . . Bi,i−1 ).
3. For every i, for every value pai of the parents PAi in V of Xi , and every
value fi = {bi , µi , σ2i } of Fi = {Bi , Mi , Σ2i },


X
bij (xj − µj ), σ 2i  .
ρ(xi |pai , fi , G) = N xi ; µi +
Xj ∈PAi
The method shown here first appeared in [Geiger and Heckerman, 1994]. We
will update parameters relative to a multivariate normal sample using Theorem
7.28, and then convert the result to a Gaussian Bayesian network. However, as
proven in Theorems 7.27 and 7.29, the prior and posterior distributions relative to a multivariate normal sample are t distributions rather than Gaussian.
So if we want our prior and updated distributions relative to the sample to be
contained in Gaussian Bayesian networks, we can only approximately represent
these distributions. We do this as follows: We develop a Gaussian Bayesian network that approximates our prior distribution (We say approximates because,
if our beliefs are represented by a multivariate normal sample, our prior distribution would have to be a t distribution.). Then we convert this network to
the corresponding multivariate normal distribution which has density function
N (x; µ, T−1 ). We approximate this distribution by a multivariate t distribution
that has density function t(x; α, µ, T), where the value of α must be assessed.
This is our prior density function. After determining our prior values of α and v,
we use Equality 7.22 to determine our prior value of β. Next we apply Theorem
7.28 to determine updated values µ∗ , α∗ ,.v∗ , and β ∗ . Then we use Equality 7.22
432
CHAPTER 7. MORE PARAMETER LEARNING
to determine the updated density function t(x; α∗ , µ∗ , T∗ ), and we approximate
this density function by N (x; µ∗ , (T∗ )−1 ). Finally, we convert this multivariate
normal distribution back to a Gaussian Bayesian network. However, when we
convert it back there is no guarantee we will obtain a network entailing the
same conditional independencies as our original network (which we assumed we
know). Therefore, for each variable Xi we convert back to a complete Gaussian
Bayesian network in which the set of parents of Xi are the same set of parents
Xi has in our original Bayesian network. For each variable Xi the resultant
updated values of bij and σi yield a density function, conditional on values of
the parents of Xi , which approximates the actual density function conditional
on values of the parents of Xi and the data. Therefore, if the conditional independencies entailed by the DAG in our original Gaussian Bayesian network are
correct (i.e. they are the ones in the actual relative frequency distribution), the
product of these conditional density functions approximates the joint density
function of the variables conditional on the data. Note that if the conditional
independencies entailed by that DAG are the ones in the relative frequency
distribution of the variables, we will have convergence to that distribution.
Note further that we never actually assess probability distributions for the
random variables Fi in an augmented Gaussian Bayesian network. Instead we
assess distributions for the random variables A and R representing unknown
mean vector and unknown precision matrix in a multivariate normal distribution. Our assumptions are actually those stated for a multivariate normal
sample, and we only use the Bayesian network to obtain prior beliefs and to
estimate the updated distribution.
We summarize our steps next. Suppose we have a multivariate normal sample D = {X(1) , X(2) , . . . X(M ) } with unassessed values of the parameters µ, v,
β, and α, Then we assess values for these parameters and approximate the
posterior density function ρX(M+1) (x(M+1) |d) as follows:
1. Construct a prior Gaussian Bayesian network (G, P ) containing for 1 ≤
i ≤ n initial values of
µi ,
σ2i ,
and
bij .
2. Convert the Gaussian Bayesian network to the corresponding multivariate
normal distribution using the algorithm in Section 7.2.3. Suppose the
density function for that distribution is N (x; µ, T−1 ).
3. Assess prior values of α and v. Recall
The parameter v is the size of the hypothetical sample upon which we
base our prior belief concerning the value of the unknown mean.
It seems reasonable to make α equal to v − 1.
4. Use Equality 7.22 to access the prior value of β. That Equality yields
β=
v(α − n + 1) −1
T .
(v + 1)
7.2. CONTINUOUS VARIABLES
433
5. Apply Theorem 7.28 to determine updated values β ∗ , α∗ , µ∗ , and v ∗ .
The components µ∗i of the updated mean vector µ∗ are the updated
unconditional means in our updated Gaussian Bayesian network.
6. Use Equality 7.22 to determine the updated value of T∗ . That Equality
yields
(v∗ + 1)
−1
β∗ .
(T∗ ) = ∗ ∗
v (α − n + 1)
7. For each variable Xi
(a) Create an ordering of the variables such that all and only the parents
of Xi in G are numbered before Xi .
−1
(b) Using the algorithm in Section 7.2.3 convert N (x(M +1) ; µ∗ , (T∗ ) )
to a Gaussian Bayesian network yielding updated values
σ∗2
i
b∗ij .
and
8. Estimate the distribution of X(M +1) by the Gaussian Bayesian network
∗
containing the DAG G and the parameter values µ∗i , σ∗2
i , and bij .
Example 7.26 Suppose we have three variables X1 , X2 , and X3 , we know that
X1 and X2 are independent, and we obtain the following data:
Case
1
2
3
4
X1
1
5
2
8
X2
2
8
4
6
X3
6
2
1
3
We apply the previous steps to learn parameter values.
1. Construct a prior Gaussian Bayesian network (G, P ). Suppose it is the
network in Figure 7.16, where all the parameters have arbitrary values.
We will see their values do not matter because we will set v = 0 to model
prior ignorance.
2. Convert the Gaussian Bayesian network to the corresponding multivariate
normal distribution using the algorithm in Section 7.2.3. We need only
use Equality 7.29 to do this. We obtain the density function N (x; µ, T−1 )
where


µ1
µ =  µ2 
µ3
and
T−1

σ21

0
=
b31 σ21
0
σ22
b32 σ22

b31 σ 21
.
b32 σ 22
2
2
2 2
2
b31 σ 1 + b32 σ2 + σ3
434
CHAPTER 7. MORE PARAMETER LEARNING
F1 2
:1
F22
:2
X1
X2
b31
b32
X3
F3 2
:3
Figure 7.16: The prior Gaussian Bayesian network for Example 7.26.
3. Assess prior values of α and v. As discussed following Theorem 7.28, to
model prior ignorance we set
v=0
α = −1.
4. Determine the prior value of β. We have
β
=
=
v(α − n + 1) −1
T
(v + 1)

σ21
0(−1 − 3 + 1) 
0
(0 + 1)
b σ2
31 1
0
σ22
b32 σ22

b31 σ21
 = 0.
b32 σ22
2 2
2
2
2
b31 σ1 + b32 σ 2 + σ3
Note that we obtained the value of β that we said we would use to model
prior ignorance (See the discussion following Theorem 7.28.). Note further
that we would have obtained the same result if we had an arc from X1 to
X2 . In this case, the independence assumption only matters when we
convert back to a Gaussian Bayesian network after learning.
5. Apply Theorem 7.28 to determine updated values β ∗ , α∗ , µ∗ , and v ∗ . As
obtained in Example 7.22,


30
18
−6
20 −10 
and
α∗ = 3,
β ∗ =  18
−6 −10 14
EXERCISES
435
and


4
µ∗ =  5 
3
v ∗ = 4.
and
So in our updated Gaussian Bayesian network
µ∗1 = 4
µ∗2 = 5
µ∗3 = 3.
6. Determine the updated value of T∗ . We have
(T∗ )−1
(v ∗ + 1)
β∗
− n + 1)


30
18
−6
4+1
 18
20 −10 
=
4(3 − 3 + 1)
−6 −10 14


37.5
22.5
−7.5
=  22.5
25
−12.5  .
−7.5 −12.5 17.5
=
v ∗ (α∗
7. For each variable Xi
(a) Create an ordering of the variables such that all and only the parents
of Xi in G are numbered before Xi .
(b) Using the algorithm in Section 7.2.3 convert N (x(M +1) ; µ∗ , (T∗ )−1 )
to a Gaussian Bayesian network yielding updated values
σ∗2
i
and
b∗ij .
For variables X1 and X3 we can use the ordering [X1 , X2 , X3 ]. To transform our multivariate normal distribution to a Gaussian Bayesian network
with this ordering, we need only use Equality 7.30. We obtain
σ∗2
1 = 37.5
b∗21 = .6
σ ∗2
2 = 11.5
b∗31 = .217
σ ∗2
3 = 10.435
b∗32 = −.696.
∗2
∗
∗
Only the values of σ∗2
1 , σ 3 , b31 , and b32 are used in our updated Gaussian
Bayesian network. To obtain the value of σ∗2
2 we use an ordering in which
X2 is numbered first. It is left as an exercise to show this value is 25.
Bernardo and Smith [1994] derive a formula for representing the updated t
distribution of X in a Bayesian network exactly.
EXERCISES
Section 7.1
436
CHAPTER 7. MORE PARAMETER LEARNING
Exercise 7.1 Find a rectangular block (not necessarily a cube) and label the
sides. Given that you are going to repeatedly throw the block, determine a
Dirichlet density function that represents your belief concerning the relative frequencies of the sides occurring and show the function. What is your probability
of each side coming up on the first throw?
Exercise 7.2 Suppose we are going to sample individuals, who have smoked
two packs of cigarettes or more daily for the past 10 years. We will determine
whether each individual’s systolic blood pressure is ≤ 100,101 − 120, 121 − 140,
141 − 160, or ≥ 161. Ascertain a Dirichlet density function that represents your
belief concerning the relative frequencies and show the function. What is your
probability of each blood pressure range for the first individual sampled?
Exercise 7.3 Prove Lemma 7.1.
Exercise 7.4 Prove Theorem 8.2.
Exercise 7.5 Prove Lemma 7.2.
Exercise 7.6 Prove Lemma 7.3.
Exercise 7.7 Prove Theorem 7.2.
Exercise 7.8 Prove Theorem 7.3.
Exercise 7.9 Prove Theorem 7.4.
Exercise 7.10 Throw the block discussed in Exercise 7.1 100 times. Compute
the probability of obtaining the data that occurs, the updated Dirichlet density
function that represent your belief concerning the relative frequencies, and the
probability of each side for the 101st throw.
Exercise 7.11 Suppose we sample 100 individuals after determining the blood
pressure ranges in Exercise 7.2, and we obtain the following results:
Blood Pressure Range
≤ 100
101 − 120
121 − 140
141 − 160
≥ 161
# of Individuals in this Range
2
15
23
25
35
Compute the probability of obtaining these data, the updated Dirichlet density
function that represent your belief concerning the relative frequencies, and the
probability of each blood pressure range for the next individual sampled.
EXERCISES
437
Exercise 7.12 Suppose F1 , F2 , . . . Fr , have a Dirichlet distribution. That is,
their density function is given by
Γ(N )
a
f1 1−1 f2a2 −1 · · · frar −1
ρ(f1 , f2 , . . . fr−1 ) = Q
r
Γ(ak )
k=1
0 ≤ fk ≤ 1,
r
X
fk = 1,
k=1
P
where a1 , a2 , . . . ar are integers ≥ 1, and N = rk=1 ak . Show that by integrating
over the remaining variables, we obtain that the marginal density function of Fk
is given by
Γ(N )
a
f k−1 (1 − fk )bk −1 ,
ρ(fk ) =
Γ(ak )Γ(bk ) k
where
bk = N − ak .
Exercise 7.13 Using the Dirichlet density function you developed in Exercise
7.2, determine 95% probability intervals for the random variables that represent
your prior belief concerning the relative frequencies of the 5 blood pressure ranges
discussed in that exercise.
Exercise 7.14 Using the posterior Dirichlet density function obtained in Exercise 7.11, determine 95% probability intervals for the random variables that
represent your posterior belief concerning the relative frequencies of the 5 blood
pressure ranges discussed in that exercise.
Exercise 7.15 Using the Dirichlet density function you developed in Exercise
7.2, determine a 95% probability square for the random variables that represents
your belief concerning the relative frequencies of the first two blood pressure
ranges discussed in that exercise.
Exercise 7.16 Using the Dirichlet density function you developed in Exercise
7.2, determine a 95% probability cube for the random variables that represent
your belief concerning the relative frequencies of the first three blood pressure
ranges discussed in that exercise.
Exercise 7.17 Prove Theorem 7.5.
Exercise 7.18 Prove Theorem 7.6.
Exercise 7.19 Prove Theorem 7.7.
Exercise 7.20 Suppose we have the augmented Bayesian network in Figure 7.6
and these data:
438
CHAPTER 7. MORE PARAMETER LEARNING
Case
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
X1
1
1
2
2
3
2
3
2
3
1
1
2
3
2
1
X2
2
3
1
2
4
2
3
1
4
1
4
1
2
3
1
Compute the following:
1. P (d)
2. ρ(f111 , f112 |d)
3. ρ(f211 , f212, f213 |d)
4. ρ(f221 , f222, f223 |d)
5. ρ(f231 , f232, f233 |d).
Show the updated augmented Bayesian network and the updated embedded Bayesian
network. For 1 ≤ k ≤ 4 determine P (X2 = k) for the 16th case.
Exercise 7.21 Does the network in Figure 7.6 have an equivalent sample size?
If so, what is it?
Exercise 7.22 Prove Theorem 7.8.
Exercise 7.23 Prove Theorem 7.9.
Exercise 7.24 Consult the references mentioned in Section 6.6 in order to extend the results in that section to the case of multinomial variables.
Section 7.2
Exercise 7.25 Establish Relation 7.3.
Exercise 7.26 Establish Relation 7.8.
EXERCISES
439
Exercise 7.27 Prove Theorem 7.14.
Exercise 7.28 Prove Theorem 7.15.
Exercise 7.29 Prove Theorem 7.16.
Exercise 7.30 At the end of Example 7.14 it was left as an exercise to show
ρR|d (r) = gamma(r; M/2, s/2). Do this.
Exercise 7.31 Obtain Equality 7.14.
Exercise 7.32 Obtain Relation 7.16.
Exercise 7.33 Obtain Equality 7.19.
Exercise 7.34 Prove Theorem 7.19.
Exercise 7.35 In Example 7.15 it was left as an exercise to show sr = s/σ2 is
distributed χ2 (M − 1). Do this.
Exercise 7.36 Prove Theorem 7.20. Hint: Show that
·
¸
¡ 2
¢
1
1
2
2
2
x − 2px1 x2 + x2
N (x1 , x2 ; 0, 1 , 0, σ , p) =
exp −
2(1 − p2 ) 1
2π (1 − p2 )1/2
·
· 2¸
¸
(x2 − px1 )2
−x1
1
1
exp
= √ exp
,
√
2
2 (1 − p2 )
2π
2π (1 − p2 )1/2
and then integrate over x2 to obtain the marginal density of X1 .
Exercise 7.37 Prove Theorem 7.21.
Exercise 7.38 Show the matrix
µ
1 0
0 1
¶
µ
1 1
1 1
¶
is positive definite.
Exercise 7.39 Show the matrix
is positive semidefinite but not positive definite.
Exercise 7.40 Show directly that if n = 1, then W ishart(v; k, 1/σ2 ) is equal
to gamma(v; k/2, 1/2σ2 ).
Exercise 7.41 Show that in the case where n = 1 the density function in Equality 7.21 is the univariate t density function which appears in Equality 7.13.
440
CHAPTER 7. MORE PARAMETER LEARNING
Exercise 7.42 Prove Theorem 7.29.
Exercise 7.43 Obtain Equality 7.29.
Exercise 7.44 Obtain the covariance matrix is Example 7.23.
Exercise 7.45 Obtain the covariance matrix is Example 7.24.
Exercise 7.46 Obtain the results in Example 7.25.
Exercise 7.47 Show the value of σ∗2
2 is 25 in Example 7.26.
Exercise 7.48 Redo Example 7.26 using a prior Gaussian Bayesian network
in which
σ22 = 1
σ 23 = 1
σ21 = 1
b21 = 0
b31 = 1
b32 = 1
µ1 = 2
µ2 = 3
µ3 = 4.
Chapter 8
Bayesian Structure
Learning
In the previous two chapters we assumed we knew the structure of the DAG
in a Bayesian network (i.e. the conditional independencies in the relative frequency distribution of a set of random variables.). Then we developed a method
for learning the parameters (estimates of the values of the relative frequencies).
Here we assume only that we have a set of random variables with an unknown
relative frequency distribution, and we develop a Bayesian method for learning
DAG structure from data. Section 8.1 develops the basic structure learning
methodology in the case of discrete variables. When a single structure is not
found to be overwhelmingly most probable, averaging over structures is sometimes more appropriate. This topic is discussed in Section 8.2. Section 8.3
concerns learning structure when there are missing data items. Next Section
8.4 discusses the problem of probabilistic model selection in a general setting,
shows that our DAG structure learning problem is an example of probabilistic
model selection, and it further shows that the model selection method we developed satisfies an important criterion (namely consistency) for a model selection
methodology. Learning structure in the case of hidden variables is the focus of
Section 8.5. Section 8.6 presents a method for learning structure in the case
of continuous variables. The terms ‘probabilistic model’ and ‘model selection’
are defined rigorously in Section 8.4. In the first three sections, by model we
always mean a candidate DAG (or DAG pattern), and we call the search for a
DAG (or DAG pattern) that best fits some data model selection.
8.1
Learning Structure: Discrete Variables
Given a repeatable experiment, whose outcome determines the state of n random
variables, in this section we assume the following:
1. The relative frequency distribution of the variables admits a faithful DAG
441
442
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
representation.
2. Given we believe the relative frequency distribution has all and only some
set of conditional independencies, our beliefs concerning the probability of
the outcome of M executions of the experiment is modeled by a multinomial Bayesian network sample with parameter (G, F) such that G entails
all and only those conditional independencies.
In Example 2.10, we showed that if a distribution has the conditional independencies
IP ({X}, {Y }|{Z}),
IP ({X}, {Y })
and only these conditional independencies, then the distribution does not admit
a faithful DAG representation. In Example 2.11, we showed that if a distribution
has the conditional independencies
IP ({L}, {F, S})
IP ({F }, {L, V })
IP ({L}, {S})
IP ({L}, {F }).
IP ({F }, {V })
and only these conditional independencies, then the distribution does not admit
a faithful DAG representation. So our first assumption above is that conditional
independencies like these are not the case for the relative frequency distribution.
The second assumption above is simply the assumption we made in the
previous two chapters when we were concerned with learning parameters.
After we develop a schema for learning structure based on these assumptions,
we show how to learn structure using the schema. Then we show how to learn
structure from a mixture of observational and experimental data. Finally, we
discuss the complexity of structure learning.
8.1.1
Schema for Learning Structure
Recall from Section 2.3.1 that Markov equivalence divides the set of all DAGs
containing the same nodes into disjoint equivalence classes, and all DAG in a
given Markov equivalence class are faithful to the same probability distributions. Recall further that we can create a graph called a DAG pattern which
represents each Markov equivalence class; and that if P admits a faithful DAG
representation, then Theorem 2.6 says there is a unique DAG pattern which is
faithful to P . Therefore, although we cannot identify a unique DAG with the
conditional independencies in P , we can identify a unique DAG pattern with
these conditional independencies. We will use GP as a random variable whose
possible values are DAG patterns gp. As far as the actual relative frequency
distribution is concerned, a DAG pattern event gp is the event that gp is faithful
to the relative frequency distribution.
In some situations we may consider DAGs events. For example, if an event
is the causal structure among the variables, then X1 → X2 represents the event
that X1 causes X2 , while X2 → X1 represents the different event that X2 causes
8.1. LEARNING STRUCTURE: DISCRETE VARIABLES
443
X1 . However, unless otherwise stated, we only consider DAG patterns events,
and as in Section 6.4.4, the notation ρ|G denotes the density function in the
augmented Bayesian network containing the DAG G. It does not entail that
the DAG G is an event.
We now have the following definition concerning learning structure:
Definition 8.1 The following constitutes a multinomial Bayesian network
structure learning schema:
1. n random variables X1 , X2 , . . . Xn with discrete joint probability distribution P ;
2. an equivalent sample size N ;
3. for each DAG pattern gp containing the n variables, a multinomial augmented Bayesian network (G, F(G) , ρ|G) with equivalent sample size N ,
where G is any member of the equivalence class represented by gp, such
that P is the probability distribution in its embedded Bayesian network.
Note that we denoted the dependence of F and ρ on G in the previous
definition whereas in the previous two chapters we did. The reason is that now
we are dealing with more than one DAG, whereas in the previous chapters the
DAG was part of our background knowledge.
Example 8.1 We develop a multinomial Bayesian network structure learning
schema containing two variables:
1. Specify two random variables X1 and X2 , each having space {1, 2}, and
assign
P (X1 = 1, X2 = 1) = 1/4
P (X1 = 1, X2 = 2) = 1/4
P (X1 = 2, X2 = 1) = 1/4
P (X1 = 2, X2 = 2) = 1/4.
2. Specify N = 4.
3. The two DAG patterns are shown in Figures 8.1 (a) and (c), and the
augmented Bayesian networks are shown in Figures 8.1 (b) and (d).
Recall G can be any element of the equivalence class represented by gp. So
the DAG in Figure 8.1 (b) could be X1 ← X2 . Note that gp1 represents no
independencies, while gp2 represents IP ({X1 }, {X2 }).
Note that, even though a Bayesian network containing the DAG X1 → X2
can contain a distribution in which X1 and X2 are independent, the event gp1
is the event they are dependent and therefore does not allow the possibility that
they are independent. This is why, when we condition on gp1 , we make F21 and
444
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
X1
X2
X1
(a) gp1
beta(f11; 2,2)
F11
(c) gp 2
beta(f21; 1,1)
beta(f22; 1,1)
F21
F22
X1
X2
beta(f11; 2,2)
beta(f21; 2,2)
F11
X2
F21
X1
(b)
X2
(d)
Figure 8.1: DAG patterns are in (a) and (c) and their respective augmented
Bayesian network structures are in (b) and (d).
F22 independent. Recall from Section 6.4.1 that we do this only when X1 and
X2 are dependent.
In general, we do not directly assign a joint probability distribution because
the number of values in the joint distribution grows exponentially with the number of variables. Rather we assign conditional distributions in all the augmented
Bayesian networks such that the probability distributions in all the embedded
Bayesian networks are the same. A common way to do this is use Theorem
7.8 to construct, for each DAG pattern gp, an augmented Bayesian network
with equivalent sample size N , whose embedded Bayesian network contains the
uniform distribution. That is, for a given DAG pattern gp we first determine a
DAG G in the equivalence class it represents. Then in the augmented Bayesian
network corresponding to G for all i, j, and k we set
aijk =
N
.
ri qi
Recall ri is the number of possible values of Xi in G, and qi is the number of
different instantiations of the parents of Xi in G. This is how the networks in
Figure 8.1 were created. Heckerman and Geiger [1995] discuss other methods
for assessing priors.
.
8.1. LEARNING STRUCTURE: DISCRETE VARIABLES
8.1.2
445
Procedure for Learning Structure
Next we show how we can learn structure using a multinomial Bayesian network
structure learning schema. We start with this definition:
Definition 8.2 The following constitutes a multinomial Bayesian network
structure learning space:
1. a multinomial Bayesian network structure learning schema containing the
variables X1 , X2 , . . . Xn ;
2. a random variable GP whose range consists of all DAG patterns containing
the n variables, and for each value gp of GP a prior probability P (gp);
3. a set D = {X(1) , X(2) , . . . X(M) } of n-dimensional random vectors such
(h)
that each Xi has the same space as Xi (See Definition 7.4.)
For each value gp of GP , D is a multinomial Bayesian network sample
of size M with parameter (G, F(G) ), where (G, F(G) ) is the multinomial
augmented Bayesian network corresponding to gp in the specification of
the schema.
Suppose we have such a space, and a set d (data) of values of the vectors in
D. Owing to Corollary 7.6,
(G)
P (d|gp) = P (d|G) =
n qY
i
Y
i=1 j=1
(G)
Γ(Nij
)
(G)
ri
Y
Γ(a(G)
ijk + sijk )
(G)
Γ(Nij
+ Mij(G) ) k=1
Γ(a(G)
ijk )
,
(8.1)
(G)
(G)
where aijk and sijk are their values in (G, F(G) , ρ|G).
A scoring criterion for a DAG (or DAG pattern) is a function which
assigns a value to each DAG (or DAG pattern) under consideration based on the
data. The expression in Equality 8.1 is called the Bayesian scoring criterion
scoreB , and is used to score both DAGs and DAG patterns. That is,
scoreB (d, gp) = scoreB (d, G) = P (d|G).
Scoring criteria are discussed in a more general way in Section 8.4.
Note that in Equality 8.1 we condition on a DAG pattern to compute the
probability that D = d, and in the previous two chapters we did not. Recall in
the previous two chapters we assumed we knew the DAG structure in all our
computations. So this structure was part of the prior background knowledge in
developing our probability space, and therefore we did not condition on it. Note
further that, due to Lemma 7.4, this conditional probability is uniquely defined
(That is, it does depend on our choice of DAGs for (G, F(G) , ρ|G).).
Given a multinomial Bayesian network structure learning space and data,
model selection consists of determining and selecting the DAG patterns with
maximum probability conditional on the data (In general, there could be more
than one maximizing DAG pattern.). The purpose of model selection is to learn
a DAG pattern along with its parameter values (a model) that can be used for
inference and decision making. Examples follow.
446
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
Example 8.2 Suppose we are doing a study concerning individuals who were
married by age 30, and we want to see if there is a correlation between graduating
college and getting divorced. We first specify the following random variables.
Variable
X1
X2
Value
1
2
1
2
When the Variable Takes this Value
Individual graduated college
Individual did not graduate college
Individual was divorced by age 50
Individual was not divorced by age 50
Next we represent our prior beliefs using the Bayesian network structure
learning schema in Example 8.1. Then the DAG pattern gp1 in Figure 8.1 (a)
represents the event that they are correlated and the DAG pattern gp2 in Figure
8.1 (b) represents the event that they are independent.
Suppose next we obtain the data d in the following table:
Case
1
2
3
4
5
6
7
8
X1
1
1
1
2
1
2
1
2
X2
1
2
1
2
1
1
1
2
Then
P (d|gp1 ) =
³
Γ(4) Γ(2+5)Γ(2+3)
Γ(4+8)
Γ(2)Γ(2)
= 7. 2150 × 10−6
P (d|gp2 ) =
³
´³
Γ(2) Γ(1+4)Γ(1+1)
Γ(2+5)
Γ(1)Γ(1)
Γ(4) Γ(2+5)Γ(2+3)
Γ(4+8)
Γ(2)Γ(2)
= 6. 746 5 × 10−6 .
´³
P (gp1 ) = P (gp2 ) = .5,
then owing to Bayes’ Theorem
=
=
Γ(2) Γ(1+1)Γ(1+2)
Γ(2+3)
Γ(1)Γ(1)
Γ(4) Γ(2+5)Γ(2+3)
Γ(4+8)
Γ(2)Γ(2)
If we assign
P (gp1 |d) =
´³
P (d|gp1 )P (gp1 )
P (d)
7. 2150 × 10−6 (.5)
P (d)
α(3. 607 5 × 10−6 )
´
´
8.1. LEARNING STRUCTURE: DISCRETE VARIABLES
P(X1=1) = 7/12
447
P(X2=1|X1=1) = 5/7
P(X2=1|X1=2) = 2/5
X1
X2
Figure 8.2: The Bayeisan network developed in Example 8.2.
and
P (d|gp1 )P (gp1 )
P (d)
6. 746 5 × 10−6 (.5)
=
.
P (d)
= α(3. 373 25 × 10−6 ),
P (gp2 |d) =
where α is a normalizing constant equal to 1/P (d). Eliminating α we have
3. 607 5 × 10−6
3. 607 5 × 10−6 + 3. 373 25 × 10−6
= .51678
P (gp1 |d) =
and
3. 373 25 × 10−6
3. 607 5 × 10−6 + 3. 373 25 × 10−6
= .48322.
P (gp2 |d) =
We select DAG pattern gp1 and conclude it is more probable that college attendance and divorce are correlated.
Furthermore, we could develop a Bayesian network, whose DAG is in the
equivalence class represented by gp1 , to do inference involving X1 and X2 . Such
a Bayesian network is shown in Figure 8.2. The parameter values in that network were learned using the technique developed in Chapter 6 (Corollary 6.7).
For the 9th case, it could be used, for example, to compute
P (X1
= 2|X2 = 1)
=
=
P (X2 = 1|X1 = 2)P (X1 = 2)
P (X2 = 1|X1 = 1)P (X1 = 1) + P (X2 = 1|X1 = 2)P (X1 = 2)
(2/5)(5/12)
= .28571.
(5/7)(7/12) + (2/5)(5/12)
Example 8.3 Suppose we are doing the same study as in Example 8.1, and we
obtain the data d in the following table:
448
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
Case
1
2
3
4
5
6
7
8
X1
1
1
1
1
2
2
2
2
X2
1
1
1
1
2
2
2
2
Then
P (d|gp1 ) =
³
Γ(4) Γ(2+4)Γ(2+4)
Γ(4+8)
Γ(2)Γ(2)
= 8. 6580 × 10−5
P (d|gp2 ) =
³
´³
Γ(2) Γ(1+4)Γ(1+0)
Γ(2+4)
Γ(1)Γ(1)
Γ(4) Γ(2+4)Γ(2+4)
Γ(4+8)
Γ(2)Γ(2)
= 4. 6851 × 10−6 .
´³
´³
Γ(2) Γ(1+0)Γ(1+4)
Γ(2+4)
Γ(1)Γ(1)
Γ(4) Γ(2+4)Γ(2+4)
Γ(4+8)
Γ(2)Γ(2)
´
´
If we assign
P (gp1 ) = P (gp2 ) = .5,
then proceeding as in Example 8.2 we obtain
P (gp1 |d) = .94866
and
P (gp2 |d) = .05134.
Notice that we become highly confident the DAG pattern is the one with the
dependency because in the data the variables are deterministically related. We
conclude that college attendance and divorce are probably correlated.
Example 8.4 Suppose we are doing the same study as in Example 8.1, and we
obtain the data d in the following table:
Case
1
2
3
4
5
6
7
8
X1
1
1
1
1
2
2
2
2
X2
1
1
2
2
1
1
2
2
8.1. LEARNING STRUCTURE: DISCRETE VARIABLES
Then
P (d|gp1 ) =
³
Γ(4) Γ(2+4)Γ(2+4)
Γ(4+8)
Γ(2)Γ(2)
= 2. 4050 × 10−6
P (d|gp2 ) =
³
´³
Γ(2) Γ(1+2)Γ(1+2)
Γ(2+4)
Γ(1)Γ(1)
Γ(4) Γ(2+4)Γ(2+4)
Γ(4+8)
Γ(2)Γ(2)
= 4. 6851 × 10−6 .
´³
´³
449
Γ(2) Γ(1+2)Γ(1+2)
Γ(2+4)
Γ(1)Γ(1)
Γ(4) Γ(2+4)Γ(2+4)
Γ(4+8)
Γ(2)Γ(2)
´
´
If we assign
P (gp1 ) = P (gp2 ) = .5,
then proceeding as in Example 8.2 we obtain
P (gp1 |d) = .33921
and
P (gp2 |d) = .66079.
Notice that we become fairly confident the DAG pattern is the one with the
independency because in the data the variables are independent. We conclude it
is more probable that college attendance and divorce are independent.
8.1.3
Learning From a Mixture of Observational and Experimental Data.
Our Bayesian scoring criterion (Equality 8.1) was derived assuming each case
obtained its value according to the same probability distribution. So we can
use it to learn structure only when all the data is observational. That is, when
no values of any variables are obtained by performing a randomized control
experiment (RCE) (See Section 1.4.1.). However, in general, we can have both
observational data and experimental data (data obtained from an RCE) on a
given set of variables. For example, in the medical domain, a great deal of
observational data is contained in routinely collected electronic medical records.
In addition, for certain variables of high clinical interest, we sometimes have data
obtained from an RCE. Cooper and Yoo [1999] developed a method for using
Equality 8.1 to score DAGs using a mixture of observational and experimental
data. We describe their method next.
First we show Equality 8.1 again:
(G)
P (d|gp) = P (d|G) =
n qY
i
Y
i=1 j=1
(G)
Γ(Nij )
(G)
(G)
ri
Y
Γ(aijk + sijk )
Γ(Nij(G) + Mij(G) ) k=1
Γ(a(G)
ijk )
.
Note that for each variable Xi , for each value of the parent set of Xi , we have
the term
(G)
(G)
(G)
ri
Y
Γ(aijk + sijk )
Γ(Nij )
.
(G)
(G)
(G)
Γ(Nij + Mij ) k=1
Γ(aijk )
450
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
When computing that term, we use all and only those cases in which the value
of Xi was obtained from observational data, regardless of how the parents of Xi
obtained their values. An example follows:
Example 8.5 Suppose we have the data d in the following table, where the
values obtained using manipulation appear bold-faced and primed:
Case
1
2
3
4
5
6
7
8
9
10
11
X1
2
2
2
1
1
20
10
2
1
2
1
X2
2
1
2
1
2
2
1
20
20
10
10
The score of DAG pattern gp1 in Figure 8.1 is as follows:
´³
´³
´
³
Γ(4) Γ(2+4)Γ(2+5)
Γ(2) Γ(1+2)Γ(1+1)
Γ(2) Γ(1+1)Γ(1+3)
P (d|gp1 ) =
Γ(4+9)
Γ(2)Γ(2)
Γ(2+3)
Γ(1)Γ(1)
Γ(2+4)
Γ(1)Γ(1)
= 4. 509 4 × 10−6 .
Note that the term for X1 is based on only 9 cases. This is because Cases 6
and 7 obtained their values of X1 via manipulation, and therefore they are not
used in the computation of that term. Note further, that the terms for X2 are
based on only 7 cases (three in which X1 has value 1 and four in which it has
value 2). This is because Cases 8, 9, 10 and 11 obtained their values of X2
via manipulation, and therefore they are not used in the computation of those
terms.
The scoring methodology just presented has been used in several algorithms
and investigations ([Tong and Koller, 2001], [Pe’er et al, 2001]).We assumed the
manipulation is deterministic. Cooper and Yoo [1999] discuss handling the
situation in which the manipulation is stochastic. Cooper [ 2000] describes
learning from a mixture of observational, experimental, and case-control (biased
sample) data.
8.1.4
Complexity of Structure Learning
When there are only a few variables, we can exhaustively compute the probability of all possible DAG patterns as was done in the previous examples. We
then select the values of gp that maximize P (d|gp) (Note that there could be
more than one maximizing pattern.) However, when the number of variables
8.2. MODEL AVERAGING
451
is not small, to find the maximizing DAG patterns by exhaustively considering all DAG patterns is computationally unfeasible. That is, Robinson [1977]
has shown the number of DAGs containing n nodes is given by the following
recurrence:
f (n) =
n
³n´
X
2i(n−i) f(n − i)
(−1)i+1
i
i=1
n>2
(8.2)
f (0) = 1
f (1) = 1.
It is left as an exercise to show f (2) = 3, f (3) = 25, f (5) = 29, 000, and
f (10) = 4.2 × 1018 . There are less DAG patterns than there are DAGs, but
this number also is forbiddingly large. Indeed, Gillispie and Pearlman [2001]
show that an asymptotic ratio of the number of DAGs to DAG patterns equal
to about 3.7 is reached when the number of nodes is only 10. Chickering [1996a]
has proven that for certain classes of prior distributions the problem of finding
the most probable DAG patterns is NP-complete.
One way to handle a problem like this is to develop heuristic search algorithms. Such algorithms are the focus of Section 9.1.
8.2
Model Averaging
Heckerman et al [1999] illustrate that when the number of variables is small and
the amount of data is large, one structure can be orders of magnitude more likely
than any other. In such cases model selection yields good results. However,
recall in Example 8.2 we had little data, we obtained P (gp1 |d) = .51678 and
P (gp2 |d) = .48322, we chose DAG pattern gp1 because it was the more probable,
and we used a Bayesian network based on this pattern to do inference for the 9th
case. Since the probabilities of the two models are so close, it seems somewhat
arbitrary to choose gp1 . So model selection does not seem appropriate. Next
we describe another approach.
Instead of choosing a single DAG pattern (model) and then using it to do
inference, we could use the law of total probability to do the inference as follows:
We perform the inference using each DAG pattern and multiply the result (a
probability value) by the posterior probably of the structure. This is called
model averaging.
Example 8.6 Recall that given the Bayesian network structure learning schema
and data discussed in Example 8.2,
P (gp1 |d) = .51678
and
P (gp2 |d) = .48322.
452
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
Case
1
2
3
4
5
X1
1
1
?
1
2
X2
1
?
1
2
?
X3
2
1
?
1
?
Table 8.1: Data on 5 cases with some data items missing
Suppose we wish to compute P (X1 = 2|X2 = 1) for the 9th trial. Since neither
DAG structure is a clear ‘winner’, we could compute this conditional probability
by ‘averaging’ over both models. To that end,
(9)
P (X1
(9)
= 2|X2
= 1, d) =
2
X
(9)
P (X1
i=1
(9)
= 2|X2
(9)
= 1, gpi , d)P (gpi |X2
= 1, d).
(8.3)
Note that we now explicitly show that this inference concerns the 9th case using
(9)
a superscript. To compute this probability, we need P (gpi |X2 = 1, d), but we
have P (gpi |d). We could either approximate the former probability by the latter
one, or we could use the technique which will be discussed in Section 8.3 to
compute it. For the sake of simplicity, we will approximate it by P (gpi |d). We
have then
P (X1(9) = 2|X2(9) = 1, d) ≈
2
X
i=1
P (X1(9) = 2|X2(9) = 1, gpi , d)P (gpi |d)
= (.28571) (.51678) + (.41667) (.48322)
= .34899.
(9)
(9)
The result that P (X1 = 2|X2 = 1, gp1 , d) = .28571 was obtained in Example
(9)
(9)
8.2. It is left as an exercise to show P (X1 = 2|X2 = 1, gp2 , d) = .41667.
Note that we obtained a significantly different conditional probability using model
averaging than that obtained using model selection in Example 8.2.
As is the case for model selection, when the number of possible structures is
large, we cannot average over all structures. In these situations we heuristically
search for high probability structures, and then we average over them. Such
techniques are discussed in Section 9.2.
8.3
Learning Structure with Missing Data
Suppose now our data set has data items missing at random as discussed in
Section 6.5. Table 8.1 shows such a data set. The straightforward way to
handle this situation is to apply the law of total probability and sum over all
the variables with missing values. That is, if D is the set of random variables
8.3. LEARNING STRUCTURE WITH MISSING DATA
453
for which we have values, d is the set of these values, and M is the set of random
variables whose values are missing, for a given DAG G,
X
scoreB (d, G) = P (d|G) =
P (d, m|G).
(8.4)
m
´T
³
is a random vector whose value
For example, if X(h) = X1(h) · · · Xn(h)
is the data for the hth case in Table 8.1, we have for the data set in that table
that
(1)
(1)
(1)
(2)
(2)
(3)
(4)
(4)
(4)
(5)
D = {X1 , X2 , X3 , X1 , X3 , X2 , X1 , X2 , X3 , X1 }
and
(2)
(3)
(3)
(5)
(5)
M = {X2 , X1 , X3 , X2 , X3 }.
We can compute each term in the sum in Equality 8.4 using Equality 8.1. Since
this sum is over an exponential number of terms relative to the number of missing data items, we can only use it when the number of missing items is not
large. To handle the case of a large number of missing items we need approximation methods. One approximation method is to use Monte Carlo techniques.
We discuss that method first. In practice, the number of calculations needed
for this method to be acceptably accurate can be quite large. Another more
efficient class of approximations uses large-sample properties of the probability
distribution. We discuss that method second.
8.3.1
Monte Carlo Methods
We will use a Monte Carlo method called Gibb’s sampling to approximate the
probability of data containing missing items. Gibb’s sampling is one variety of
an approximation method called Markov Chain Monte Carlo (MCMC).
So first we review MCMC.
Review of Markov Chains and MCMC
First we review Markov chains; then we review MCMC; finally we show the
MCMC method called Gibb’s sampling.
Markov Chains This exposition is only for the purpose of review. If you are
unfamiliar with Markov chains, you should consult a complete introduction as
can be found in [Feller, 1968]. We start with the definition:
Definition 8.3 A Markov chain consists of the following:
1. A set of outcomes (states) e1 , e2 , . . . .
2. For each pair of states ei and ej a transition probability pij such that
X
pij = 1.
j
454
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
e1
e2
e2
e3
e2
e3
e1
e2
e3
e1
e1
e2
e2
e3
e1
e2
e3
e2
e1
e2
e1
e2
e3
e3
Figure 8.3: An urn model of a Markov chain.
3. A sequence of trials (random variables) E (1) , E (2) , . . . such that the outcome of each trial is one of the states, and
P (E (h+1) = ej |E (h) = ei ) = pij .
To completely specify a probability space we need define initial probabilities
P (E (0) = ej ) = pj , but these probabilities are not necessary to our theory and
will not be discussed further.
Example 8.7 Any Markov chain can be represented by an urn model. One
such model is shown in Figure 8.3. The Markov chain is obtained by choosing
an initial urn according to some probability distribution, picking a ball at random
from that urn, moving to the urn indicated on the ball chosen, picking a ball at
random from the new urn, and so on.
The transition probabilities pij are arranged in a matrix of transition probabilities as follows:


p11 p12 p13 · · ·
 p21 p22 p23 · · · 


P =  p31 p32 p33 · · ·  .


..
..
..
..
.
.
.
.
This matrix is called the transition matrix for the chain.
Example 8.8 For the Markov chain determined by the urns in Figure 8.3 the
transition matrix is


1/6 1/2 1/3
P =  2/9 4/9 1/3  .
1/2 1/3 1/6
A Markov chain is called finite if it has a finite number of states. Clearly
(n)
the chain represented by the urns in Figure 8.3 is finite. We denote by pij
the probability of a transition from ei to ej in exactly n trials. This is,
8.3. LEARNING STRUCTURE WITH MISSING DATA
455
p(n)
ij is the conditional probability of entering ej at the nth trial given the initial
state is ei . We say ej is reachable from ei if there exists an n ≥ 0 such that
(n)
pij > 0. A Markov chain is called irreducible if every state is reachable from
every other state.
Example 8.9 Clearly, if pij > 0 for every i and j, the chain is irreducible.
The state ei has period t > 1 if p(n)
ii = 0 unless n = mt for some integer m,
and t is the largest integer with this property. Such a state is called periodic.
A state is aperiodic if no such t > 1 exists.
Example 8.10 Clearly, if pii > 0, ei is aperiodic.
(n)
We denote by fij the probability that starting from ei the first entry to ej
occurs at the nth trial. Furthermore, we let
fij =
∞
X
(n)
fij .
n=1
Clearly, fij ≤ 1. When fij = 1, we call Pij (n) ≡ fij(n) the distribution of
the first passage for ej starting at ei . In particular, when fii = 1, we call
(n)
Pi (n) ≡ fii the distribution of the recurrence times for ei , and we define
the mean recurrence time for ei to be
µi =
∞
X
(n)
nfii .
n=1
The state ei is called persistent if fii = 1 and transient if fii < 1. A persistent
state ei is called null if its mean recurrence time µi = ∞ and otherwise it is
called non-null.
Example 8.11 It can be shown that every state in a finite irreducible chain is
persistent (See [Ash, 1970].), and that every persistent state in a finite chain is
non-null (See [Feller, 1968].). Therefore every state in a finite irreducible chain
is persistent and non-null.
An aperiodic persistent non-null state is called ergodic. A Markov chain is
called ergodic if all its states are ergodic.
Example 8.12 Owing to Examples 8.9, 8.10, and 8.11, if in a finite chain we
have pij > 0 for every i and j, the chain is an irreducible ergodic chain.
We have the following theorem concerning irreducible ergodic chains:
Theorem 8.1 In an irreducible ergodic chain the limits
(n)
rj = lim pij
n→∞
(8.5)
456
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
exist and are independent of the initial state ei . Furthermore, rj > 0,
X
rj = 1,
(8.6)
X
(8.7)
j
rj =
ri pij ,
i
and
rj =
1
,
µj
where µj is the mean recurrence time of ej .
The probability distribution
P (E = ej ) ≡ rj
is called the stationary distribution of the Markov chain.
Conversely, suppose a chain is irreducible and aperiodic with transition matrix P, and there exists numbers rj ≥ 0 satisfying Equalities 8.6 and 8.7. Then
the chain is ergodic, and the rj s are given by Equality 8.5.
Proof. The proof can be found in [Feller, 1968].
We can write Equality 8.7 in the matrix/vector form
rT = rT P.
(8.8)
That is,
¡
r1
r2
r3
···
¢
=
¡
r1
r2
r3
···

¢



p11
p21
p31
..
.
Example 8.13 Suppose we have the Markov chain
Figure 8.3. Then

1/6
¢ ¡
¢
¡
r1 r2 r3 = r1 r2 r3  2/9
1/2
p12
p22
p32
..
.
p13
p23
p33
..
.
···
···
···
..
.



.

determined by the urns in
1/2
4/9
1/3

1/3
1/3  .
1/6
(8.9)
Solving the system of equations determined by Equalities 8.6 and 8.9, we obtain
¢ ¡
¢
¡
r1 r2 r3 = 2/7 3/7 2/7 .
This means for n large the probabilities of being in states e1 , e2 , and e3 are
respectively about 2/7, 3/7, and 2/7 regardless of the initial state.
8.3. LEARNING STRUCTURE WITH MISSING DATA
457
MCMC Again our coverage is cursory. See [Hastings, 1970] for a more thorough introduction.
Suppose we have a finite set of states {e1 , e2 , . . . es }, and a probability distribution P (E = ej ) ≡ rj defined on the states such that rj > 0 for all j. Suppose
further we have a function f defined on the states, and we wish to estimate
I=
s
X
f (ej )rj .
j=1
We can obtain an estimate as follows.
Given we have
¢ a Markov chain with transi¡
tion matrix P such that rT = r1 r2 r3 · · · is its stationary distribution,
we simulate the chain for trials 1, 2, ...M . Then if ki is the index of the state
occupied at trial i, and
M
X
f(eki )
,
(8.10)
I0 =
M
i=1
the ergodic theorem says that I 0 → I with probability 1 (See [Tierney, 1996].).
So we can estimate I by I 0 . This approximation method is called Markov
chain Monte Carlo. To obtain more rapid convergence, in practice a burnin number of iterations is used so that the probability of being in each state is
approximately given by the stationary distribution. The sum in Expression 8.10
is then obtained over all iterations past the burn-in time. Methods for choosing
a burn-in time and the number of iterations to use after burn-in are discussed
in [Gilks et al, 1996].
It is not hard to see why the approximation converges. After a sufficient
burn-in time, the chain will be in state ej about rj fraction of the time. So if
we do M iterations after burn in, we would have
M
X
i=1
f (eki )/M ≈
s
X
f (ej )rj M
j=1
M
=
s
X
f (ej )rj .
j=1
To apply this method for a given distribution r, we need to construct a
Markov chain with transition matrix P such that r is its stationary distribution.
Next we show two ways for doing this.
Metropolis-Hastings Method Owing to Theorem 8.1, we see from Equality 8.8 that we need only find an irreducible aperiodic chain such that its transition matrix P satisfies
(8.11)
rT = rT P.
It is not hard to see that if we determine values pij such that for all i and j
ri pij = rj pji
(8.12)
the resultant P satisfies Equality 8.11. Towards determining such values, let
Q be the transition matrix of an arbitrary Markov chain whose states are the
458
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
members of our given finite set of states {e1 , e2 , . . . es }, and let

sij

qij 6= 0, qji 6= 0


 1 + ri qij
rj qji
αij =
,




0
qij = 0 or qji = 0
(8.13)
where sij is a symmetric function of i and j chosen so that 0 ≤ αij ≤ 1 for all i
and j. We then take
pij
pii
= αij qij
X
= 1−
pij .
i 6= j
(8.14)
j6=i
It is straightforward to show that the resultant values of pij satisfy Equality
8.12. The irreducibility of P must be checked in each application.
Hastings [1970] suggests the following way of choosing s: If qij and qji are
both nonzero, set

ri qij
rj qji

1+
≥1



rj qji
ri qij
.
(8.15)
sij =

rj qji
rj qji


≤1
 1+
ri qij
ri qij
Given this choice, we have


1







rj qji
αij =

ri qij







0
qij 6= 0, qji 6= 0,
rj qji
≥1
ri qij
qij 6= 0, qji 6= 0,
rj qji
≤1 .
ri qij
(8.16)
qij = 0 or qji = 0
If we make Q symmetric (That is, qij = qji for all i and j.), we have the method
devised by Metropolis et al (1953). In this case

1
qij 6= 0, rj ≥ ri





(8.17)
rj /ri
qij 6= 0, rj ≤ ri .
αij =





0
qij = 0
Note that with this choice if Q is irreducible so is P.
¡
¢
Example 8.14 Suppose rT = 1/8 3/8 1/2 . Choose Q symmetric as
follows:


1/3 1/3 1/3
Q =  1/3 1/3 1/3  .
1/3 1/3 1/3
8.3. LEARNING STRUCTURE WITH MISSING DATA
459
Choose s according to Equality 8.15 so that α has the values in Equality 8.17.
We then have


1
1
1
1 .
α =  1/3 1
1/4 3/4 1
Using Equality 8.14 we have

1/3 1/3
P =  1/9 5/9
1/12 1/4
Notice that
rT P =
=
as it should.
¡
¡
1/8 3/8
1/2
1/8 3/8
1/2
¢
¢

1/3
1/3  .
2/3

1/3 1/3
 1/9 5/9
1/12 1/4
= rT

1/3
1/3 
2/3
Once we have constructed matrices Q and α as discussed above, we can
conduct the simulation as follows:
1. Given the state occupied at the kth trial is ei , choose a state using the
probability distribution given by the ith row of Q. Suppose that state is
ej .
2. Choose the state occupied at the (k + 1)st trial to be ej with probability
αij and to be ei with probability 1 − αij .
In this way, when state ei is the current state, ej will be chosen qij fraction of
the time in Step (1), and of those times ej will be chosen αij fraction of the
time in Step (2). So overall ej will be chosen αij qij = pij fraction of the time
(See Equality 8.14.), which is what we want.
Gibb’s Sampling Method Next we show another method for creating
a Markov chain whose stationary distribution is a particular distribution. The
method is called Gibb’s sampling, and it concerns the case where we have n
random variables X1 , X2 , . . . Xn and a joint probability distribution P of the
¡
¢T
variables (as in a Bayesian network). If we let X = X1 · · · Xn
, we
want to approximate
X
f (x)P (x).
x
To approximate this sum using MCMC, we need create a Markov chain whose
set of states is all possible values of X, and whose stationary distribution is
P (x). We do this as follows: The transition probability in our chain for going
from state x0 to x00 is defined to be the product of these conditional probabilities:
460
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
P (x001 |x02 , x03 , . . . x0n )
0
P (x002 |x001 , x03 , . . . xn )
..
.
P (x00k |x001 , . . . x00k−1 , x0k+1 . . . x0n )
..
.
00 00
P (xn |x1 , . . . , x00n−1 , x00n ).
We can implement these transition probabilities by choosing the event in each
trial using n steps as follows. If we let pk (x; x̂) denote the transition probability
from x to x̂ in the kth step, we set
pk (x; x̂) =
½
P (x̂k |x̂1 , . . . x̂k−1 , x̂k+1 . . . x̂n )
0
x̂j = xj for all j 6= k
otherwise.
That is, we do the following for the hth trial:
(h)
Pick x1
(h)
(h−1)
, x3
(h)
(h−1)
using the distribution P (x1 |x2
(h−1)
, . . . x(h−1)
).
n
, . . . x(h−1)
).
Pick x2 using the distribution P (x2 |x1 , x3
n
..
.
(h)
(h)
(h)
(h−1)
).
Pick xk using the distribution P (xk |x1 , . . . xk−1 , xk+1 . . . x(h−1)
n
..
.
(h)
(h)
(h−1)
).
Pick x(h)
n using the distribution P (xn |x1 , . . . , xn−1 , xn
(h)
Notice that in the kth step, all variables except xk are unchanged, and the new
(h)
value of xk is drawn from its distribution conditional on the current values of
all the other variables.
As long as all conditional probabilities are nonzero, the chain is irreducible.
Next we verify that P (x) is the stationary distribution for the chain. If we let
p(x; x̂) denote the transition probability from x to x̂ in each trial, we need show
P (x̂) =
X
P (x)p(x; x̂).
(8.18)
x
It is not hard to see that it suffices to show Equality 8.18 holds for each each
step of each trial. To that end, for the kth step we have
8.3. LEARNING STRUCTURE WITH MISSING DATA
X
461
P (x)pk (x; x̂)
x
=
X
P (x1 , . . . xn )pk (x1 , . . . xn ; x̂1 , . . . x̂n )
x1 ,...xn
=
X
xk
P (x̂1 , . . . x̂k−1 , xk , x̂k+1 . . . x̂n )P (x̂k |x̂1 , . . . x̂k−1 , x̂k+1 . . . x̂n )
= P (x̂k |x̂1 , . . . x̂k−1 , x̂k+1 . . . x̂n )
X
P (x̂1 , . . . x̂k−1 , xk , x̂k+1 . . . x̂n )
xk
= P (x̂k |x̂1 , . . . x̂k−1 , x̂k+1 . . . x̂n )P (x̂1 , . . . x̂k−1 , x̂k+1 . . . x̂n )
= P (x̂1 , . . . x̂k−1 , x̂k , x̂k+1 . . . x̂n )
= P (x̂).
The second step follows because pk (x; x̂) = 0 unless x̂j = xj for all j 6= k.
See [Geman and Geman, 1984] for more on Gibb’s sampling.
Learning with Missing Data Using Gibb’s Sampling
The Gibb’s sampling approach we use is called the Candidate method (See
[Chib, 1995].). The approach proceeds as follows: Let d be the set of values of
the variables for which we have values. By Bayes’ Theorem we have
P (d|G) =
P (d|f̌ (G) , G)ρ(f̌ (G) |G)
,
ρ(f̌ (G) |d, G)
(8.19)
where f̌ (G) is an arbitrary assignment of values to the parameters in G. To
approximate P (d|G) we choose some value of f̌ (G) , evaluate the numerator in
Equality 8.19 exactly, and approximate the denominator using Gibb’s sampling.
For the denominator, we have
X
ρ(f̌ (G) |d, m, G)P (m|d, G)
ρ(f̌ (G) |d, G) =
m
where M is the set of variables which have missing values.
To approximate this sum using Gibb’s sampling we do the following:
1. Initialize the state of the unobserved variables to arbitrary values yielding
a complete data set d1 .
(h)
2. Choose some unobserved variable Xi
(h)
Xi using
0(h)
P (xi
arbitrarily and obtain a value of
0(h)
(h)
P (x , d1 − {x̌i }|G)
(h)
|d1 − {x̌i }, G) = X i (h)
(h)
P (xi , d1 − {x̌i }|G)
(h)
xi
462
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
(h)
where x̌(h)
is the value of Xi in d1 , and the sum is over all values in
i
(h)
the space of Xi . The terms in the numerator and denominator can be
computed using Equality 8.1.
3. Repeat step (2) for all the other unobserved variables, where the complete
data set used in the (k + 1)st iteration contains the values obtained in the
previous k iterations.
This will yield a new complete data set d2 .
4. Iterate the previous two steps some number R times where the complete
data set from the the jth iteration is used in the (j + 1)st iteration. In
this manner R complete data sets will be generated. For each complete
data set dj compute
ρ(f̌ (G) |dj , G)
using Corollary 7.7.
5. Approximate
ρ(f̌ (G) |d, G) ≈
PR
j=1
ρ(f̌ (G) |dj , G)
.
R
Although the Candidate method can be applied with any value of f̌ (G) of
the parameters, some assignments lead to faster convergence. Chickering and
Heckerman [1997] discuss methods for choosing the value.
8.3.2
Large-Sample Approximations
Although Gibb’s sampling is accurate, the amount of computer time needed
to achieve accuracy can be quite large. An alternative approach is the use
of large-sample approximations. Large-sample approximations require only a
single computation and choose the correct model in the limit. So they can
be used when the size of the data set is large. We discuss four large-sample
approximations next.
Before doing this, we need to further discuss the MAP and ML values of
the parameter set. Recall in Section 6.5 we introduced these values in a context
which was specific to binomial Bayesian networks and in which we needn’t specify a DAG because the DAG was part of our background knowledge. We now
provide notation appropriate to this chapter. Given a multinomial augmented
Bayesian network (G, F(G) , ρ|G), the MAP value f̃ (G) of f (G) is the value that
maximizes ρ(f (G) |d, G), and the maximum likelihood (ML) value f̂ (G) of f (G)
is the value such that P (d|f (G) , G) is a maximum. In the case of missing data
items, Algorithm 6.1 (EM-MAP-determination) can be used to obtain approximations to these values. That is, if we apply Algorithm 6.1 and we obtain the
values s0G
ijk , then
(G)
0(G)
aijk + sijk
³
´
(G)
0(G)
a
+
s
k=1
ijk
ijk
(G)
f˜ijk ≈ P
ri
8.3. LEARNING STRUCTURE WITH MISSING DATA
463
Similarly, if we modify Algorithm 6.1 to estimate the ML value (as discussed
after the algorithm) and we obtain the values s0G
ijk , then
0(G)
sijk
(G)
fˆijk ≈ Pr
.
0(G)
i
k=1 sijk
In the case of missing data items, these approximations are the ones which
would be used to compute the MAP and ML values in the formulas we develop
next.
The Laplace Approximation
First we derive the Laplace Approximation. This approximation is based on the
assumptions that ρ(f (G) |d, G) has a unique MAP value f̂ (G) and its logarithm
allows a Taylor Series expansion about f̂ (G) . These conditions hold for multinomial augmented Bayesian networks. As we shall see in Section 8.5.3, they do
not hold when we consider DAGs with hidden variables.
For the sake of notational simplicity, we do not show the dependence on G
in this derivation. We have
Z
P (d) = P (d|f)ρ(f)df.
(8.20)
Towards obtaining an approximation of this integral, let
g(f) = ln (P (d|f)ρ(f)) .
Owing to Bayes’ Theorem
g(f) = ln (αρ(f|d)) ,
where α is a normalizing constant, which means g(f) achieves a maximum at the
MAP value f̃. Our derivation proceeds by taking the Taylor Series expansion of
g(f) about f̃. To write this expansion we denote f as a random vector f . That
is, f is the random vector whose components are the members of the set f. We
denote f̃ by f̃ . Discarding terms past the second derivative, this expansion is
T
T
1
g(f ) ≈ g(f̃ ) + (f − f̃ ) g0 (f̃ ) + (f − f̃ ) g00 (f̃ )(f − f̃ ),
2
where g 0 (f ) is the vector of first partial derivatives of g(f ) evaluated with respect to every parameter fijk , and g 00 (f ) is the Hessian matrix of second
partial derivatives of g(f ) evaluated with respect to every pair of parameters
(fijk , fi0 j 0 k0 ). That is,
g0 (f ) =
³
∂g(f )
∂f111
∂g(f )
∂f112
···
´
,
464
and
CHAPTER 8. BAYESIAN STRUCTURE LEARNING



g00 (f ) = 

∂ 2 g(f )
∂f111 ∂f111
∂ 2 g(f )
∂f111 ∂f112
∂ 2 g(f )
∂f112 ∂f111
..
..
.
..
.
.
···



··· .

..
.
Now g0 (f̃ ) = 0 because g(f ) achieves a maximum at f̃ , which means its derivative
is equal to zero at that point. Therefore,
T
1
g(f ) ≈ g(f̃ ) + (f − f̃ ) g00 (f̃ )(f − f̃ ).
2
(8.21)
By ≈ we mean ‘about equal to’. The approximation in Equality 8.21 is guaranteed to be good only if f is close to f̃ . However, when the size of the data set is
large, the value of P (d|f) declines fast as one moves away from f̃ , which means
only values of f close to f̃ contribute much to the integral in Equality 8.20. This
argument is formalized in [Tierney and Kadane, 1986].
Owing to Equality 8.21, we have
Z
P (d) =
P (d|f)ρ(f)df
Z
=
exp (g(f)) df
¶
µ
³
´Z
T 00
1
(f − f̃ ) g (f̃ )(f − f̃ ) df
(8.22)
≈ exp g(f̃ )
exp
2
Recognizing that the expression inside the integral in Equality 8.22 is proportional to a multivariate normal density function (See Section 7.2.2.), we obtain
that
³
´
³
´
−1/2
= exp P (d|f̃ )ρ(f̃ ) 2π d/2 |A|−1/2 , (8.23)
P (d) ≈ exp g(f̃ ) 2πd/2 |A|
00
where
Pn A = −g (f̃ ), and d is the number of parameters in the network, which is
i=1 qi (ri − 1). Recall ri is the number of states of Xi and qi is the number of
possible instantiations of the parents PAi of Xi . In general, d is the dimension
of the model given data d in the region of f̃ . If we do not make the assumptions
leading to Equality 8.23, d is not necessarily the number of parameters in the
network. We discuss such a case in Section 8.5.3.We have then that
³
´
³
´ d
1
(8.24)
ln (P (d)) ≈ ln P (d|f̃ ) + ln ρ(f̃ ) + ln(2π) − ln |A| .
2
2
The expression in Equality 8.24 is called the Laplace approximation or
Laplace score. Reverting back to showing the dependence on G and denoting
the parameter set again as a set, we have that this approximation is given by
´
³
´ d
³
1
Laplace (d, G) ≡ ln P (d|f̃ (G) , G) +ln ρ(f̃ (G) |G) + ln(2π)− ln |A| . (8.25)
2
2
8.3. LEARNING STRUCTURE WITH MISSING DATA
465
To select a model using this approximation, we choose a DAG (and thereby
the DAG pattern representing the equivalence class to which the DAG belongs)
which maximizes Laplace (d, G). The value of P (d|f̃ (G) , G) can be computed
using a Bayesian network inference algorithm.
We say an approximation method for learning a DAG model is asymptotically correct if, for M (the sample size) sufficiently large, the DAG selected
by the approximation method is one that maximizes P (d|G). Kass et al [1988]
show that under certain regularity conditions
|ln (P (d|G)) − Laplace (d, G)| ∈ O(1/M ),
(8.26)
where M is the sample size and the constant depends on G. For the sake of
simplicity we have not shown the dependence of d on M . It is not hard to see
that Relation 8.26 implies the Laplace approximation is asymptotically correct.
The BIC Approximation
It is computationally costly to determine the value of |A| in the Laplace approximation. A more efficient but less accurate approximation can be obtained by
retaining only those terms in Equality 8.25 that are not bounded as M increases.
Furthermore, as M approaches ∞, the determinant |A| approaches a constant
times M d , and the MAP value f̃ (G) approaches the ML value f̂ (G) . Retaining
only the unbounded terms, replacing |A| by M d , and using f̂ (G) instead of f̃ (G) ,
we obtain the Bayesian information criterion(BIC) approximation or
BIC score, which is
´ d
³
BIC (d, G) ≡ ln P (d|f̂ (G) , G) − ln M,
2
Schwarz [1978] first derived the BIC approximation. It is not hard to see that
Relation 8.26 implies
|ln (P (d|G)) − BIC (d, G)| ∈ O(1).
(8.27)
It is possible to show the following two conditions hold for a multinomial
Bayesian network structure learning space (Note that we are now showing the
dependence of d on M .):
1. If we assign proper prior distributions to the parameters, for every DAG
G we have
lim P (dM |G) = 0.
M →∞
2. If GM is a DAG which maximizes P (dM |G), then for every G not in the
same Markov equivalence class as GM ,
P (dM |G)
= 0.
M→∞ P (dM |GM )
lim
466
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
It is left as an exercise to show that these two facts along with Relation 8.27
imply the BIC approximation is asymptotically correct.
The BIC approximation is intuitively appealing because it contains 1) a term
which shows how well the model predicts the data when the parameter set is
equal to its ML value; and 2) a term which punishes for model complexity. Another nice feature of the BIC is that it does not depend on the prior distribution
of the parameters, which means there is no need to assess one.
The MLED Score
Recall that to handle missing values when learning parameter values we used
Algorithm 6.1 (EM-MAP-determination) to estimate the MAP value f̃ of the
parameter set f. The fact that the MAP value maximizes the posterior distribution of the parameters suggests approximating the probability of d using a
fictitious data set d0 that is consistent with the MAP value. That is, we use the
number of occurrences obtained in Algorithm 6.1 as the number of occurrences
in an imaginary data set d0 to obtain an approximation. We have then that
(G)
0
M LED (d, G) ≡ P (d |G) =
n qY
i
Y
i=1 j=1
Γ(Nij(G) )
0(G)
ri
Y
Γ(a(G)
ijk + sijk )
Γ(Nij(G) + Mij(G) ) k=1
Γ(a(G)
ijk )
,
0(G)
where the values of sijk are obtained using Algorithm 6.1. We call this approximation the marginal likelihood of the expected data (MLED) score.
Note that we do not call MLED an approximation because it computes the
probability of fictitious data set d0 , and d0 could be substantially larger than d,
which means it could have a much smaller probability. So MLED could only be
used to select a DAG pattern, not to approximate the probability of data given
a DAG pattern.
Using MLED, we select a DAG pattern which maximizes P (d0 |G). However,
as discussed in [Chickering and Heckerman, 1996], a problem with MLED is
that it is not asymptotically correct. Next we develop an adjustment to it that
is asymptotically correct.
The Cheeseman-Stutz Approximation
The Cheeseman-Stutz approximation or CS score, which was originally
proposed in [Cheeseman and Stutz, 1995], is given by
³
´
³
´
CS(d, G) ≡ ln (P (d0 |G)) − ln P (d0 |f̂ (G) , G) + ln P (d|f̂ (G) , G) ,
where d0 is the imaginary data set introduced in the previous subsection. The
value of P (d0 |f̂ (G) , G) can readily be computed using Lemma 6.11. The formula
in that lemma extends immediately to multinomial Bayesian networks.
Next we show the CS approximation is asymptotically correct. We have
8.3. LEARNING STRUCTURE WITH MISSING DATA
467
³
´
³
´
CS(d, G) ≡ ln (P (d0 |G)) − ln P (d0 |f̂ (G) , G) + ln P (d|f̂ (G) , G)
¸ ·
¸
·
d
d
0
0
= ln (P (d |G)) − BIC (d , G) + ln M + BIC (d, G) + ln M
2
2
0
0
= ln (P (d |G)) − BIC (d , G) + BIC (d, G) .
So
ln (P (d|G)) − CS(d, G)
= [ln (P (d|G)) − BIC (d, G)] + [BIC (d0 , G) − ln (P (d0 |G))]
(8.28)
Relation 8.27 and Equality 8.28 imply
|ln (P (d|G)) − CS (d, G)| ∈ O(1).
which means the CS approximation is asymptotically correct.
The CS approximation is intuitively appealing for the following reason. If
we use this approximation to actually estimate the value of ln(P (d|G)), then
our estimate of P (d|G) is given by
#
"
P (d0 |G)
P (d|f̂ (G) , G).
P (d|G) ≈
P (d0 |f̂ (G) , G)
That is, we approximate the probability of the data by its probability given the
ML value of the parameter set, but with an adjustment based on d0 .
A Comparison of the Approximations
Chickering and Heckerman [1997] compared the accuracy and computer times
of the approximations methods. Their analysis is very detailed, and you should
consult the original source for a complete understanding of their results. Briefly,
they used a model to generate data, and then compared the results of the
Laplace, BIC, and CS approximations to those of the Gibb’s sampling Candidate method. That is, this latter method was considered the gold standard.
Furthermore, they used both MAP and ML values in the BIC and CS (We
presented them with ML values).
First, they used the Laplace, BIC, and CS approximations as approximations of the probability of the data given candidate models. They compared
these results to the probabilities obtained using the Candidate method. They
found that the CS approximation was more accurate with the MAP values,
but the BIC approximation was more accurate with the ML values. Furthermore, with the MAP values, the CS approximation was about as accurate as
the Laplace approximation, and both were significantly more accurate than the
BIC approximation. This result is not unexpected since the BIC approximation
includes a constant term.
468
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
In the case of model selection, we are really only concerned with how well
the method selects the correct model. Chickering and Heckerman [1997] also
compared the models selected by the approximation methods with that selected
by the Candidate method. They found the CS and Laplace approximations both
selected models which were very close to that selected by the Candidate method,
and the BIC approximation did somewhat worse. Again the CS approximation
performed better with the MAP values.
As to time usage, the order is what we would expect. If we consider the
time used by the EM algorithm separately, the order of time usage in increasing
order is as follows: 1) BIC/CS; 2) EM; 3) Laplace; 4) Candidate. Furthermore,
the time usage increased significantly with model dimension for the Laplace
algorithm, whereas it hardly increased for the BIC, CS, and EM algorithms. As
the dimension went from 130 to 780, the time usage for the Laplace algorithm
increased over 10 fold to over 100 seconds and approached that of the Candidate
algorithm. On the other hand, the time usage for the BIC and CS algorithms
stayed close to 1 second, and the time usage for the EM algorithm stayed close
to 10 seconds.
Given the above, of the approximation methods presented here, the CS approximation seems to be the method of choice. Chickering and Heckerman
[1996,1997] discuss other approximations based on the Laplace approximation,
which fared about as well as the CS approximation in their studies.
8.4
Probabilistic Model Selection
The structure learning problem discussed in Section 8.1 is an example of a more
general problem called probabilistic model selection. After defining ‘probabilistic model’, we discuss the general problem of model selection. Finally we show
that the selection method we developed satisfies an important criterion (namely
consistency) for a model selection methodology.
8.4.1
Probabilistic Models
A probabilistic model M for a set of random variables V is a set of joint probability distributions of the variables. Ordinarily, each joint probability distribution in a model is obtained by assigning values to the members of a parameter
set F which is part of the model. If probability distribution P is a member of
model M, we say P is included in M. If the probability distributions in a
model are obtained by assignments of values to the members of a parameters
set F, this means there is some assignment of values to the parameters that
yields the probability distribution. Note that this definition of ‘included’ is a
generalization of the one in Section 2.3.2. An example of a probabilistic model
follows.
Example 8.15 Suppose we are going to toss a die and a coin, neither of which
are known to be fair. Let X be a random variables whose value is the outcome
8.4. PROBABILISTIC MODEL SELECTION
469
of the die toss, and let Y be a random variable whose value is the outcome
of the coin toss. Then the space of X is {1, 2, 3, 4, 5, 6} and the space of Y is
{heads, tails}. The following is a probabilistic model M for the joint probability
distribution of X and Y :
P6
1. F = {f11 , f12 , f13 , f14 , f15 , f16 , f21 , f22 }, 0 ≤ fij ≤ 1,
j=1 f1j = 1,
P2
f
=
1.
j=1 2j
2. For each permissible combination of the parameters in F, obtain a member
of M as follows:
P (X = i, Y = heads) = f1i f21
P (X = i, Y = tails) = f1i f22.
Any probability distribution of X and Y for which X and Y are independent
is included in M; any probability distribution of X and Y for which X and Y
are not independent is not included M.
A Bayesian network model (also called a DAG model) consists of a
DAG G =(V, E), where V is a set of random variables, and a parameter set F
whose members determine conditional probability distributions for the DAGs,
such that for every permissible assignment of values to the members of F, the
joint probability distribution of V is given by the product of these conditional
distributions and this joint probability distribution satisfies the Markov condition with the DAG. Theorem 1.5 shows that if F determines discrete probability
distributions, the product of the conditional distributions will satisfy the Markov
condition. After this theorem, we noted the result also holds if F determines
Gaussian distributions. For simplicity, we ordinarily denote a Bayesian network
model using only G (i.e. we do not show F.). Note that an augmented Bayesian
network (Definition 6.8) is based on a Bayesian network model. That is, given
an augmented Bayesian network (G, F(G) , ρ|G), (G, F(G) ) is a Bayesian network
model. We say the augmented Bayesian network contains the Bayesian network
model.
Example 8.16 Bayesian network models appear in Figures 8.4 (a) and (b).
The probability distribution contained in the Bayesian network in Figure 8.4 (c)
is included in both models, whereas the one in the Bayesian network in Figure
8.4 (d) is included only in the model in Figure 8.4 (b).
A set of models, each of which is for the same set of random variables, is
called a class of models.
Example 8.17 The set of Bayesian networks models contained in the set of all
multinomial augmented Bayesian networks containing the same variables is a
class of models. We call this class a multinomial Bayesian network model
class. Figure 8.4 shows models from the class when V = {X1 , X2 , X3 }, X1 and
X3 are binary, and X2 has space size three.
470
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
f111
f211
f212
f221
f222
f311
f321
f331
f111
f221
f222
f311
f321
f331
f341
f351
f361
X1
X2
X3
X1
X2
X3
f211
f212
(a)
(b)
P(X3=1|X1=1,X2=1) = .2
P(X3=1|X1=1,X2=2) = .7
P(X2=1|X1=1) = .2
P(X2=2|X1=1) = .5
P(X1=1) = .2
P(X2=1|X1=2) = .4
P(X2=2|X1=2) = .1
X1
X2
(c)
P(X2=1|X1=1) = .2
P(X2=2|X1=1) = .5
P(X3=1|X2=1) = .2
P(X3=1|X2=2) = .7
P(X3=1|X2=3) = .6
X2
P(X1=1) = .2
X1
P(X2=1|X1=2) = .4
P(X2=2|X1=2) = .1
X2
P(X3=1|X1=1,X2=3) = .6
P(X3=1|X1=2,X2=1) = .9
P(X3=1|X1=2,X2=2) = .4
P(X3=1|X1=2,X2=3) = .3
X3
(d)
Figure 8.4: Bayesian network models appear in (a) and (b). The probability
distribution in the Bayesian network in (c) is included in both models, whereas
the one in (d) is included only in the model in (b).
A conditional independency common to all probability distributions included
in model M is said to be in M. We have the following theorem:
Theorem 8.2 In the case of a Bayesian network model G, the set of conditional
independencies in model G is the set of all conditional independencies entailed
by d-separation in DAG G.
Proof. The proof follows immediately from Theorems 2.1.
Model M1 is distributionally included in model M2 (denoted M1 ≤D
M2 ) if every distribution included in M1 is included in M2 . If M1 is distributionally included in M2 and there exists a probability distribution which
is included in M2 and not in M1 , we say M1 strictly distributionally included in M2 (denoted M1 <D M2 ). If M1 is distributionally included in M2
and no such probability distribution exists, we say they are distributionally
equivalent (denoted M1 ≈D M2 ). Model M1 is independence included in
model M2 (denoted M1 ≤I M2 ) if every conditional independency in M2 is
in M1 . If M1 is both distributionally included and independence included in
M2 , we simply say M1 is included in M2 (denoted M1 ≤ M2 ). Definitions
analogous to those for distributional inclusion hold for strictly independence
included, independence equivalent, strictly included, and equivalent.
8.4. PROBABILISTIC MODEL SELECTION
471
Theorem 8.3 For a multinomial Bayesian network class, G1 is distributionally
included in G2 if and only if G1 is independence included in G2 .
Proof. Suppose G1 is independence included in G2 . Let P be a distribution
included in G1 . Then every conditional independency in G1 is in P , which
means every conditional independency in G2 is a conditional independency in
P . Owing to Theorem 8.2, this means every d-separation in G2 is a conditional
independency in P , which means P satisfies the Markov condition with G2 . The
fact that P is included in G2 now follows from Theorem 1.4.
In the other direction, suppose G1 is not independence included in G2 . Then
there exists some conditional independence (d-separation) I in G2 which is not
in G1 . Let P be a probability distribution included in G1 which is faithful to G1
(As mentioned following Example 2.9, almost all assignments of values to the
conditional distributions will yield such a P .). This P does not have conditional
independence I, which means P is not included in G2 . Therefore, G1 is not
distributionally included in G2 .
Owing to the previous theorem, when we discuss models from a multinomial
Bayesian network class we speak only of inclusion. Note that in this case models
are equivalent if and only if their DAGs are Markov equivalent.
Example 8.18 Suppose our models are from a multinomial Bayesian network
class. Model G2 = X1 → X2 → X3 is strictly included in the model G1 , whose
DAG contains the three variables but only has the edge X1 → X2 because the
latter model contains more conditional independencies. Therefore, G1 < G2 .
Example 8.19 Suppose our models are from a multinomial Bayesian network
class. Model X1 → X2 → X3 is equivalent to model X1 ← X2 ← X3 since models in this class are equivalent if and only if their DAGs are Markov equivalence.
Given some class of models, if M2 includes probability distribution P and
there exists no M1 in the class such that M1 includes P and M1 <D M2 ,
then M2 is called a distributionally inclusion optimal map. Analogous
definitions hold for independence inclusion optimal map and inclusion
optimal map.
Suppose we have a multinomial Bayesian network class of models for a set of
random variables V. Then if a probability distribution P of V admits a faithful
DAG representation, DAG G is faithful to P if and only if model G is an inclusion
optimal map of P . However, if DAG G satisfies the minimality condition with P
(See Definition 2.11.), model G is not necessarily an inclusion optimal map. For
example, the DAG G in Figure 2.18 (c) satisfies the minimality condition with
the DAG referenced in that figure, but model G is not an inclusion optimal map
since the model containing the DAG in Figure 2.18 (a) is strictly included in
it. On the other hand, if model G is an inclusion optimal map of P , then DAG
G satisfies the minimality condition with P . Section 8.4.3 discusses a model
G (namely the one containing the DAG in Figure 8.6) which is an inclusion
optimal map of P but DAG G is not faithful to P .
472
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
Given some class of models, if M2 includes probability distribution P and
there exists no M1 in the class such that M1 includes P and M1 has smaller
dimension than M2 , then M2 is called a parameter optimal map of P . In the
case of Bayesian network models (DAG models), the dimension of the model
is the number of parameters in the models. However, as we shall in Section
8.5.3, this is not always the case. Theorem 8.6 will show that in the case of a
multinomial Bayesian network class, a parameter optimal map is an inclusion
optimal map.
8.4.2
The Model Selection Problem
In general, the problem of model selection is to find a concise model which,
based on a random sample of observations from the population that determines
a relative frequency distribution (See Section 4.2.1.), includes an approximation of the relative frequency distribution. As previously done, we use d to
represent the set of values (data) of the sample. To perform model selection,
we develop a scoring function score (called a scoring criterion) which assigns a
value score(d, M) to each model under consideration based on the data. We
have the following definition concerning scoring criteria:
Definition 8.4 Let dM be a set of M values (data) of a set of random variables,
score be a scoring criterion over some class of models for the random variables,
and PM be the joint distribution determined by the data dM . We say score is
consistent for the class of models if the following two properties hold:
1. For M sufficiently large, if M1 includes PM and M2 does not, then
score(dM , M1 ) > score(dM , M2 ).
2. For M sufficiently large, if M1 and M2 both include PM and M1 has
smaller dimension than M2 , then
score(dM , M1 ) > score(dM , M2 ).
We call the distribution determined by the data the generative distribution. Henceforth, we use that terminology.
If the data set is sufficiently large, a consistent scoring criterion chooses a
parameter optimal map of the generative distribution. This parameter optimal
map is attractive for the following reason: If the set of values of the random
variables is a random sample from an actual relative frequency distribution
and we accept the von Mises theory (See Section 4.2.1.), then as the size of the
data set becomes large the generative distribution approaches the actual relative
frequency distribution. Therefore, a parameter optimal map, of the generative
distribution, will in the limit be a most parsimonious model that includes the
actual relative frequency distribution.
8.4. PROBABILISTIC MODEL SELECTION
8.4.3
473
Using the Bayesian Scoring Criterion for Model Selection
First we show the Bayesian scoring criterion is consistent. Then we discuss using
it when the faithfulness assumption is not warranted.
Consistency of Bayesian Scoring
If the actual relative frequency distribution admits a faithful DAG representation, our goal is to find a DAG (and its corresponding DAG pattern) which is
faithful to that distribution. If it does not, we would want to find a DAG G such
that model G is a parameter optimal independence map of that distribution. If
we accept the von Mises theory (See Section 4.2.1.), then a consistent scoring
criterion (See Definition 8.4.) will accomplish the latter task when the size of
the data set is large. Next we show the Bayesian scoring criterion is consistent.
After that, we show that in the case of DAGs a consistent scoring criterion finds
a faithful DAG if one exists.
Lemma 8.1 In the case of a multinomial Bayesian network class, the BIC
scoring criterion (See Section 8.3.2.) is consistent for scoring DAGs.
Proof. Haughton [1988] shows that this lemma holds for a class consisting of
curved exponential models. Geiger at al [1998] show a multinomial Bayesian
network class is such a class.
Theorem 8.4 In the case of a multinomial Bayesian network class, the Bayesian
scoring criterion scoreB (d, G) = P (d|G) is consistent for scoring DAGs.
Proof. The Bayesian scoring criterion scores a model G in a multinomial
Bayesian network class by computing P (d|G) using a multinomial augmented
Bayesian network containing G. In Section 8.3.2 we showed that for multinomial
augmented Bayesian networks, the BIC score is asymptotically correct, which
means for M (the sample size) sufficiently large, the model selected by the BIC
score is one that maximizes P (d|G). The proof now follows from the previous
lemma.
Before proceeding, we need the definitions and lemmas that follow.
Definition 8.5 We say edge X → Y is covered in DAG G if X and Y have
the same parents in G except X is not a parent of itself.
Definition 8.6 If we reverse a covered edge in a DAG, we call it a covered
edge reversal.
Clearly, if we perform a covered edge reversal on a DAG G we obtain a DAG
in the same Markov equivalence class as G.
Theorem 8.5 Suppose G1 and G2 are Bayesian network models such that
G1 ≤I G2 . Let r be the number of links in G2 that have opposite orientation in G1 , and let m be the number of links in G2 that do not exist in G1 in
474
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
either orientation. There exists a sequence of r + 2m distinct operations to G1 ,
where each operation is either an edge addition or a covered edge reversal, such
that
1. after each operation G1 is a DAG and G1 ≤I G2 ;
2. after all the operations G1 = G2 .
Proof. The proof can be found in [Chickering, 2002].
Definition 8.7 Size Equivalence holds for a class of Bayesian network models if models containing Markov equivalent DAGs have the same number of parameters.
It is not hard to see that size equivalence holds for a multinomial Bayesian
network class.
Theorem 8.6 Given a class of Bayesian network models for which size equivalence holds, a parameter optimal map of a probability distribution P is an
independence inclusion optimal map of P .
Proof. Let G2 be a parameter optimal map of P . If G2 is not an independence
inclusion optimal map of P , there is some model G1 which includes P and
G1 <I G2 .
Owing to Theorem 8.5, we can transform G1 to G2 with a sequence of edge
additions and covered edge reversals. If there are no edge additions, G1 and
G2 are in the same Markov equivalence class, which means G1 ≮I G2 . So
there must be at least one edge addition, which strictly increases the size of the
model. Since we assumed size equivalence, covered edge reversals leave the size
of the model unchanged. We conclude the model containing G2 contains more
parameters than the model containing G1 , which contradicts the fact that G2 is
a parameter optimal map of P .
The converse of the preceding theorem does not hold. That is, an independence inclusion optimal map is not necessarily a parameter optimal map.
Section 8.4.3 presents an example illustrating this.
Corollary 8.1 Given a multinomial Bayesian network class, if a model is a
parameter optimal map of P and P admits a faithful DAG representation, then
the DAG in the model is faithful to P .
Proof. Clearly, size equivalence holds for this class. Therefore, owing to the
previous theorem, a parameter optimal map of P is an independence inclusion
optimal map of P . It is not hard to see that for this class, if P admits a faithful
DAG representation then, if model G is an independence optimal inclusion map
of P , DAG G must be faithful to P .
Theorem 8.7 In the case of a multinomial Bayesian network class, for M
(the sample size) sufficiently large, the Bayesian scoring criterion chooses an
inclusion optimal map of the generative distribution P . If P admits a faithful
DAG representation, it chooses a model whose DAG is faithful to P .
Proof. The proof follows Theorems 8.3, 8.4, 8.6, and Corollary 8.1.
8.4. PROBABILISTIC MODEL SELECTION
N
475
L
F
T
C
Figure 8.5: We assume the probability distribution of job requring night plane
travel (N ), lung cancer (L), tuberculosis (T ), fatigue (F ), and a postive chest
X-ray (C) is faithful to this DAG.
Results When the Faithfulness Assumption is not Warranted
Recall from Section 8.1 that our structure learning methodology assumes the
relative frequency distribution of the variables admits a faithful DAG representation. Suppose we have the following random variables:
Variable
T
N
L
F
C
Value
t1
t2
n1
n2
l1
l2
f1
f2
c1
c2
When the Variable Takes this Value
Patient has tuberculosis
Patient does not have tuberculosis
Patient’s job requires night plane travel
Patient job does not require night plane travel
Patient has lung cancer
Patient does not have lung cancer
Patient is fatigued
Patient is not fatigued
Patient has a positive chest X-ray
Patient has a negative chest X-ray
Suppose further that lung cancer and the job requiring night plane travel each
cause fatigue, lung cancer and tuberculosis each cause a positive chest X-ray,
there are no other causal relationships among the variables, and there are no
hidden common causes. Then, due to the argument in Section 2.6, we would
expect the relative frequency distribution of the five variables to be faithful to
the DAG in Figure 8.5. Assume this is the case. Then owing to the result in
Example 2.11, the marginal distribution of N , F , C, and T is not be faithful
to any DAG. Assume next that we are observing only these four variables and
we obtain data on them. That is, assume we know nothing about lung cancer,
indeed we have not even identified it as a feature of humans. Then the assumption of faithfulness is not valid. Suppose we score models using the Bayesian
scoring criterion (Equality 8.1). Owing to Theorems 8.4 and 8.7, if the data set
is sufficiently large, a parameter optimal independence map, which is also an inclusion optimal map, of the generative distribution P will be chosen. Assuming
that distribution is the same as the actual relative frequency distribution, the
476
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
N
T
F
C
N
T
F
(a)
C
(b)
Figure 8.6: If P is faithful to the DAG in Figure 8.5, both DAG patterns
represent inclusion optimal models.
DAG in the model will be one from the Markov equivalence class represented
by the DAG pattern in either Figure 8.6 (a) or Figure 8.6 (b). Say it is a model
containing a DAG G from the equivalence class represented in Figure 8.6 (a).
DAG G does not entail IP ({N}, {C}), which holds for the actual relative frequency distribution. Nevertheless, since P is included in model G, DAG G is an
independence map of P , which means we can use DAG G in a Bayesian network
containing the variables. We can then use our inference algorithms for Bayesian
networks to do inference with the variables.
You may ask what determines whether a model corresponding to Figure 8.6
(a) or one corresponding to Figure 8.6 (b) is chosen. It depends on the size
of the models. For example, suppose all variables are binary except F , which
has space size three. Then a model containing a DAG in the equivalence class
represented in Figure 8.6 (a) has smaller dimension, and therefore such a model
will be chosen. If all variables are binary, the dimension is the same regardless
of the equivalence class and therefore the Bayesian score is the same.
8.5
Hidden Variable DAG Models
Next we discuss DAG models containing hidden variables. A hidden variable
DAG model is a DAG model containing a DAG G =(V ∪ H, E), where V =
{X1 , X2 , . . . Xn } and H = {H1 , H2 , . . . Hk } are disjoint sets of random variables.
The variables in V are called observable variables, while the variables in H
are called hidden variables. In practice we obtain data only on the variables
in V. By using hidden variables, we can obtain more models for V than we could
if we considered only DAG models containing DAGs of the form G =(V, E).
After discussing hidden variable DAG models in which the DAG entails
more conditional independencies than any DAG containing only the observables,
we present hidden variable DAG models in which the DAG entails the same
conditional independencies as one containing only the observables. Next we
discuss computing the dimension of a hidden variable DAG model. Then we
illustrate how changing the space sizes of the hidden variables changes the model.
Finally, we address efficient methods for scoring hidden variable DAG models.
8.5. HIDDEN VARIABLE DAG MODELS
N
477
H
F
T
C
Figure 8.7: A DAG containing a hidden variable.
8.5.1
Models Containing More Conditional Independencies than DAG Models
Recall in Section 8.4.3 we showed that if, P is faithful to the DAG in Figure
8.5, the Bayesian scoring criterion will choose a model whose DAG is in one
of the equivalence classes represented in Figure 8.6. The DAG in the model
does not entail all the conditional independencies in P . However, if we consider
only DAGs containing the four observable variables, this is the best we can
do because the probability distribution of these four variables does not admit
a faithful DAG representation. Alternatively, we could also consider a hidden
variable DAG model GH containing the DAG in Figure 8.7. The variable H is a
hidden variable because it is not one of the four variables on which we have data
(Notice that we shade nodes representing hidden variables.). This variable does
not represent lung cancer (Recall we are assuming we have not even identified
lung cancer as a feature of humans.). Rather it is a hypothesized variable about
which we have no knowledge. We could give H any size space we wish. For the
current discussion, assume it is binary. Mathematically, this is really a case of
missing data items. We simply have all data items missing for H. So we can
compute P (d|GH ) in the same way as we computed the probability of data with
missing data items in Equality 8.4. That is,
M
scoreB (d, GH ) = P (d|GH ) =
2
X
i=1
P (di |GH ),
(8.29)
where M is the size of the sample, and di is data on the five variables, N , F , C,
T , and H, with the values of N , F , C, and T being their actual ones and those
of H ranging over all of their 2M possibilities.
For example, suppose we have the data d in the following actual table:
478
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
Case
1
2
3
4
5
6
7
8
9
N
n1
n1
n2
n2
n1
n2
n2
n2
n2
F
f1
f2
f1
f2
f1
f1
f1
f1
f1
C
c2
c1
c2
c2
c1
c2
c1
c1
c2
T
t2
t2
t2
t1
t1
t2
t1
t2
t1
Then d1 and an arbitrary intermediate di are respectively the data in the following tables:
d1
Case
1
2
3
4
5
6
7
8
9
N
n1
n1
n2
n2
n1
n2
n2
n2
n2
F
f1
f2
f1
f2
f1
f1
f1
f1
f1
C
c2
c1
c2
c2
c1
c2
c1
c1
c2
T
t2
t2
t2
t1
t1
t2
t1
t2
t1
H
h1
h1
h1
h1
h1
h1
h1
h1
h1
Case
1
2
3
4
di
5
6
7
8
9
N
n1
n1
n2
n2
n1
n2
n2
n2
n2
F
f1
f2
f1
f2
f1
f1
f1
f1
f1
C
c2
c1
c2
c2
c1
c2
c1
c1
c2
T
t2
t2
t2
t1
t1
t2
t1
t2
t1
H
h1
h2
h2
h1
h2
h1
h2
h2
h1
Given we score all the possible DAG models containing only the variables
N , F , C, and T using the standard Bayesian scoring criterion, and we score
the hidden variable DAG model GH we’ve discussed using Equality 8.29, you
may ask which model should win when the sample size is large. Theorem 8.4
is not applicable because that theorem assumes the class consists of curved
exponential models, and Geiger et al [1998] show a hidden variable DAG model
G =(V ∪ H, E) is not a curved exponential model for V. Rather it is a stratified
exponential model. For hidden variable DAG models, Meek [1997] sketched a
proof showing the Bayesian scoring criterion sastifies the first requirement in a
consistent scoring criterion. That is, he illustrated that for M sufficiently large
if the hidden variable model GH includes PM and model G does not, then the
Bayesian scoring criterion will score GH higher than G. Furthermore, Rusakov
and Geiger [2002] proved that the BIC and Bayesian scoring criterion satisfy
both requirements for consistency (i.e. they show BIC and the Bayesian scoring
criterion are consistent.) in the case of naive hidden variable DAG models,
which are models as follows: There is a single hidden variable H, all observables
are children of H, and there are no edges between any observables. Model GH
is not a naive hidden variable DAG model. However, it seems the Bayesian
scoring criterion is consistent in the case of more general hidden variable DAG
models. If it consistent in the case of model GH and if the relative frequency
distribution is faithful to the DAG in Figure 8.5, then GH should be chosen
8.5. HIDDEN VARIABLE DAG MODELS
1
1
2
479
2
1
2
3
2
3
1
2
3
2
Figure 8.8: Value and shape are not independent, but they are conditionally
independent give color.
when the data set is large. It is left an an exercise to test this conjecture by
generating data and scoring the models.
A problem with using this technique lies in identifying which hidden variable DAG models to consider. That is, why should we suspect that the joint
distribution of N, F , C, and T does not admit a faithful DAG representation,
and, even if we did suspect this, why should we choose the correct hidden variable DAG model? Chapter 10 presents an algorithm for determining whether a
probability distribution admits a faithful DAG representation, and it develops a
method for discovering hidden variables when this is not the case. Furthermore,
that chapter discusses applying the theory to causal learning, and it shows that
in many applications the variable H in Figure 8.7 can be considered a hidden
common cause of F and C.
8.5.2
Models Containing the Same Conditional Independencies as DAG Models
Recall Example 2.14 showed it is possible to embed a distribution P faithfully
in two DAGs, and yet P is included in only one of the DAGs. Owing to this
fact, data can sometimes distinguish a hidden variable DAG models from a
DAG model containing no hidden variables, even when both models contain the
same conditional independencies. First we illustrate this. Then we discuss an
application to learning causal influences.
Distinguishing a Hidden Variable Model From a DAG Model
When a hidden variable DAG model and a DAG model, which has no hidden
variables, contain the same conditional independencies, the hidden variables are
somewhat obscure entities. To impart intuition for them we develop a lengthy
urn example. Please bear with us as the outcome should be worth the effort.
Suppose we have an urn containing the objects in Figure 8.8. Let random
variables V, S,and C be defined as follows:
480
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
P(c1) = 1/2
P(c2) = 1/2
C
S
V
P(s1) = 1/2
P(s2) = 1/3
P(s3) = 1/6
V
S
P(v1|c1) = 1/2
P(v2|c1) = 1/2
P(v3|c1) = 0
P(s1|c1) = 2/3
P(s2|c1) = 1/3
P(s3|c1) = 0
P(v1|c2) = 0
P(v2|c2) = 1/2
P(v3|c2) = 1/2
P(s1|c2) = 1/3
P(s2|c2) = 1/3
P(s3|c2) = 1/3
P(v1|s1) = 1/3
P(v2|s1) = 1/2
P(v3|s1) = 1/6
P(v1|s2) = 1/4
P(v2|s2) = 1/2
P(v3|s2) = 1/4
P(v1|s3) = 0
P(v2|s3) = 1/2
P(v3|s3) = 1/2
(b)
(a)
Figure 8.9: The probability distribution of V and S obtained by sampling from
the objects in Figure 8.8 is contained in both Bayesian networks.
Variable
V
S
C
Value
v1
v2
v3
s1
s2
s3
c1
c2
Outcomes Mapped to this Value
All objects containing a ‘1’
All objects containing a ‘2’
All objects containing a ‘3’
All square objects
All circular objects
All arrow objects
All black objects
All white objects
Suppose further we sample with replacement from the urn. Assuming the generative distribution is the same as the distribution obtained by applying the
principle of indifference to the objects in the urn, clearly we have
qIP (V, S)
and
IP (V, S|C),
where P denotes the generative distribution. It is not hard to see then that the
probability distribution of V , S, and C is contained in the Bayesian network in
Figure 8.9 (a), and the probability distribution of only V and S is contained in
the Bayesian networks in both Figure 8.9 (a) and Figure 8.9 (b).
Suppose next we have an urn containing the objects in Figure 8.10. Let
random variables V and S be defined as before. If we sample from this urn and
the generative distribution is same as that obtained by applying the principle of
indifference to the objects, this distribution is contained in the Bayesian network
8.5. HIDDEN VARIABLE DAG MODELS
1
2
3
2
481
1
3
3
1
1
2
2
3
Figure 8.10: There is no apparent division of the objects which renders value
and shape independent.
S
P(s1) = 3/11
P(s2) = 4/11
P(s3) = 4/11
V
P(v1|s1) = 1/3
P(v2|s1) = 1/3
P(v3|s1) = 1/3
P(v1|s2) = 1/4
P(v2|s2) = 1/4
P(v3|s2) = 1/2
P(v1|s3) = 1/4
P(v2|s3) = 1/2
P(v3|s3) = 1/4
Figure 8.11: The probability distribution of V and S obtained by sampling from
the objects in Figure 8.10 is contained in this Bayesian network.
in Figure 8.11. Could the distribution also be contained in a Bayesian network
like the one in Figure 8.9 (a)? That is, is there some division of the objects into
two groups that renders V and S independent in each group? Note that the
coloring of the objects in Figure 8.8 was only for emphasis. It is the fact that
they could be divided into two groups such that V and S are independent in
each group that enabled us to represent the probability distribution of V and S
using the Bayesian network in Figure 8.9 (a).
Towards answering the question just posed, consider the two models in Figure 8.12. The Bayesian network in Figure 8.9 (a) is obtained from the model in
Figure 8.12 (a) by assigning values to the parameters in the model. The only
difference is that we labeled the root C instead of H in the DAG in Figure 8.9
(a) because at that time we were identifying color. The Bayesian networks in
Figure 8.9 (a) and Figure 8.11 are obtained from the model in Figure 8.12 (b)
by assigning values to the parameters in the model. The question posed in the
previous paragraph can now be stated as follows: Is the probability distribution
in the Bayesian network in Figure 8.11 included in the hidden variable DAG
482
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
f111
H
V
S
f211
f212
f311
f312
f221
f222
f321
f322
(a)
S
V
f111
f112
f211
f212
f221
f222
f231
f232
(b)
Figure 8.12: A hidden variable DAG model for V and S is in (a) and a DAG
model for V and S is in (b).
model in Figure 8.12 (a)? It is very unlikely for the following reason. Although
the model in Figure 8.12 (a) has more parameters than the one in Figure 8.12
(b), some of the parameters are redundant. As a result, it effectively has fewer
parameters and its dimension is smaller. This is discussed much more in Section
8.5.3. As a result of its dimension being smaller, it includes far less distributions. Specifically, every distribution included in the model in Figure 8.12 (a)
is clearly included in the model in Figure 8.12 (b). However, if we consider the
space consisting of the set of all possible permissible assignments to the parameters in the model in Figure 8.12 (b), the subset, whose members yield joint
distributions included in the model in Figure 8.12 (a), has Lebesgue measure
zero. In summary, if we let GH be the model in Figure 8.12 (a) and G be the
model in 8.12 (b), it is that case that GH <D G.
Suppose now we randomly select either model GH or model G, we randomly
assign permissible values to the parameters, and we generate a very large amount
of data on V and S using the resultant Bayesian network. You are then given
the data and have the task of deciding which model we selected. To accomplish
this, you could score each model using the Bayesian scoring criterion and select
the one with the higher score. Recall from Section 8.5.1 that the Bayesian
scoring criterion is consistent in the case of naive hidden variable DAG models.
Clearly, model GH is a naive hidden variable model. So the Bayesian scoring
criterion is consistent when choosing between models GH and G. Therefore, for
a large data set, it will choose model G over GH if and only if the generative
distribution is not included in model GH . If we generated the data using model
GH , then the generative distribution in the case of a large amount of data should
be included in GH , and the Bayesian scoring criterion will correctly choose GH
over G. If we generated the data using model G, almost for certain the generative
8.5. HIDDEN VARIABLE DAG MODELS
Case
1
2
3
4
5
6
7
8
9
10
11
12
Sex
female
female
female
female
female
female
male
male
male
male
male
male
Height (inches)
64
64
64
64
68
68
64
64
68
68
70
70
483
Wage ($)
30, 000
30, 000
40, 000
40, 000
30, 000
40, 000
40, 000
50, 000
40, 000
50, 000
40, 000
50, 000
Table 8.2: Possible data on sex, height, and wage
distribution in the case of a large amount of data is not included GH , and the
Bayesian scoring criterion almost for certain will correctly choose G over GH .
Note that whenever you discovered a ‘hidden variable’ exists in the previous
experiment, you were really discovering there is a division of the objects into
two groups such that V and S are independent in each group.
Application to Causal Learning
The value of urn problems is that they shed light on interpreting results obtained when we model situations in the external world. Next we discuss such a
situation.
Recall Example 1.17 in which we had the data in Table 8.2. In Exercise 1.12,
we showed that if Sex, Height, and W age are random variables representing
sex, height, and wage respectively, then, if this small sample is indicative of the
probabilistic relationships P among the variables in some population, we have
IP (Height, W age|Sex).
Note that the distribution determined by the objects in Figure 8.8 is the same
as the distribution in Table 8.2 if we make the following associations:
black/f emale
arrow/70
white/male
1/30, 000
circle/64
2/40, 000
square/68
3/50, 000
Now suppose we have the data in Table 8.2, but we are not able to detect the
sex of the individuals. Sex is then a hidden variable in that it is a property of
the entities being investigated that renders two measured variables independent.
Owing to the discussion in the previous section, if we obtain a large amount of
data on height and wage reflecting the distribution in Table 8.2, the Bayesian
scoring criterion should choose the hidden variable DAG model GH in Figure
484
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
8.12 (a) over the DAG model G in Figure 8.12 (b), and we obtain evidence for
the existence of a hidden common cause of height and wage.
There are several reasons we only say that we obtain evidence for the existence of a hidden common cause. First of all, the DAGs X → H → Y and
X ← H ← Y are Markov equivalent to X ← H → Y . So the division of the data
into subsets which render Height and W age independent in each subset could
be due to an intermediate cause. Furthermore, in real applications features
like height and wage are actually continuous, and we create discrete variables
by imposing cutoff points in the data. For example, we may classify height as
68 ≤ height < 69, 69 ≤ height < 70, etc., and wage as 30, 000 ≤ wage < 40, 000,
40, 000 ≤ wage < 50, 000, etc. With these cutoffs points we may obtain a division of the data that render height and wage independent, whereas with other
cutoff points we may not.
The same cautionary note holds concerning the conclusion of the absence
of a hidden clause. Suppose model G were chosen over model GH . A different
choice of cutoff points may result in GH being chosen. Furthermore, if there is
a hidden common cause, it may be better modeled with a hidden variable that
has a larger space. Clearly, if we increase the space size of the hidden variable
sufficiently, GH will have the same dimension as G, and the data will not be
able to distinguish the two models.
8.5.3
Dimension of Hidden Variable DAG Models
Recall in Section 8.5.2 we mentioned that although the model in Figure 8.12 (a)
has more parameters than the one in Figure 8.12 (b), some of the parameters
are redundant. As a result, it effectively has fewer parameters and its dimension
is smaller. Next we discuss calculating the dimension of hidden variable DAG
models.
Let d be the dimension (size) of the model given data d in the region of the
MAP (or ML) value of the parameters. Recall that in Equality 8.23 d equals the
number of parameters in the model. That equality was obtained by assuming the
expression inside the integral in Equality 8.22 is proportional to a multivariate
normal density function. This assumption requires that there is a unique MAP
(and therefore ML) value of the parameters, and the probability of the data is
peaked at that value. This assumption holds for multinomial Bayesian network
models. We show next that it does not hold for hidden variable DAG models,
and therefore we cannot consider the number of parameters in such a model the
dimension of the model.
Consider the hidden variable DAG model in Figure 8.13 (a) and the trivial
DAG model in Figure 8.13 (b). Clearly, they include the same distributions of
X, and the dimension of the trivial DAG model is 1. We will show the dimension
of the hidden variable DAG model is also 1. Let f ={f111 , f211 , f221 } and fX be
a parameter representing the probability that X equals its first value. Then
fX = f211 f111 + f221 (1 − f111 ).
8.5. HIDDEN VARIABLE DAG MODELS
485
f111
H
f111
X
X
f211
f221
(a)
(b)
Figure 8.13: The effective dimension of the model in (a) is the same as the
dimension of the model in (b).
The probability of data on X is uniquely determined by the value of fX . That
is, if we let d be a set of M values (data) of X, s be the number of time X
equals its first value, and t be the number of times X equals its second value,
then
s
(1 − fX )t .
P (d|f) = P (d|fX ) = fX
The unique maximum likelihood value of f is
s
fˆX =
.
M
Any values of the parameters in the network such
s
= f211 f111 + f221 (1 − f111 )
M
will yield this value. So the probability of the data does not reach a peak at a
unique value of f; rather it reaches a ridge at a set of values. In this sense the
model has only one non-redundant parameter, and the dimension of the model
is 1.
Geiger et al [1996] discuss this matter more formally, and they show that,
if we determine the dimension d of the model as just illustrated, the scores discussed in Section 8.3.2 approximate the Bayesian score in the case of hidden
variable DAG models. Furthermore, they compute the dimension of some interesting hidden variable DAG models. The following table shows the dimension
of naive hidden variable DAG models when all variables are binary:
486
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
Number (n)
of Observables
Dimension (d) of
Hidden Variable DAG Model
1
2
3≤n≤7
n>7
1
3
2n + 1
2n ≤ d ≤ 2n + 1
For n > 7 they only obtained the bounds shown. Note when n is 1 or 2,
the dimension of the hidden variable DAG model is less than the number of
parameters in the model, when 3 ≤ n ≤ 7 its dimension is the same as the
number of parameters, and for n > 7 its dimension is bounded above by the
number of parameters. Note further that when n is 1, 2, or 3 the dimension of
the hidden variable DAG model is the same as the dimension of the complete
DAG model, and when n ≥ 4 it is smaller. Therefore, owing to the fact that the
Bayesian scoring criterion is consistent in the case of naive hidden variable DAG
models (discussed in Section 8.5.1), using that criterion we can distinguish the
models from data when n ≥ 4.
Let’s discuss the naive hidden variable DAG model in which H is binary
and there are two non-binary observables. Let r be space size of both observables. If r ≥ 4, the number of parameters in the hidden variable DAG model
is less than the number in the complete DAG model; so clearly its dimension is
smaller. It is possible to show the dimension is smaller even when r = 3 (See
[Kocka and Zhang, 2002].).
Finally, consider the hidden variable DAG model X → Y ← H → Z ← W ,
where H is the hidden variable. If all variables are binary, the number of
parameters in the model is 11. However, Geiger et al [1996] show the dimension
is only 9. They showed further that if the observables are binary, and H has
space size 3 or 4 the dimension 10, while if H has space size 5 the dimension is 11.
The dimension could never exceed 12 regardless of the space size of H, because
we can remove H from the model to create the DAG model X → Y → Z ← W
with X → W also, and this model has dimension 12.
8.5.4
Number of Models and Hidden Variables
At the end of the last section, we discussed varying the space size of the hidden
variable, while leaving the number of states of the observable fixed. In the case
of hidden variable DAG models, a DAG containing observables with fixed space
sizes, can be contained in different models because we can assign different space
sizes to a hidden variable. An example is AutoClass, which was developed by
Cheeseman and Stutz [1995].
Autoclass is a classification program for unsupervised learning of clusters.
The cluster learning problem is as follows: Given a collection of unclassified entities and features of those entities, organize those entities into classes that in
some sense maximize the similarity of the features of the entities in the same
class. For example, we may want to create classes of observed creatures. Autoclass models this problem using the hidden variable DAG model in Figure
8.5. HIDDEN VARIABLE DAG MODELS
487
H
D1
D2
D3
C1
C2
C3
C4
C5
C6
Figure 8.14: An example of a hidden variable DAG model used in Autoclass.
8.14. In that figure, the hidden variable is discrete, and it is possible values
correspond to the underlying classes of entities. The model assumes the features represented by discrete variables (in the figure D1 , D2 , and D3 ), and sets
of features represented by continuous variables (in the figure {C1 , C2 , C3 , C4 }
and {C5 , C6 }) are mutually independent given H. Given a data set containing
values of the features, Autoclass search over variants of this model, including
the number of possible values of the hidden variable, and it selects a variant so
as to approximately maximize the posterior probability of the variant.
The comparison studies discussed in Section 8.3.2 were performed using this
model with all variables being discrete.
8.5.5
Efficient Model Scoring
In the case of hidden variable DAG models the determination of scoreB (d, GH )
requires an exponential number of calculations. First we develop a more efficient
way to do this calculation in certain cases. Then we discuss approximating the
score.
A More Efficient Calculation
Recall that in the case of binary variables Equality 8.29 gives the Bayesian score
as follows:
2M
X
P (di |GH ),
(8.30)
scoreB (d, GH ) = P (d|GH ) =
i=1
where M is the size of the sample. Clearly, this method has exponential time
complexity in terms of M. Next we show how to do this calculation more
efficiently.
488
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
One Hidden Variable Suppose GH is S ← H → V where H is hidden, all
variables are binary, we have the data d in the following table, and we wish to
score GH based on these data:
Case
1
2
3
4
5
6
7
8
9
S
s1
s1
s2
s2
s1
s2
s2
s2
s2
V
v1
v2
v1
v2
v1
v1
v1
v1
v1
Consider the di s, represented by the following tables, which would appear in
the sum in Equality 8.30:
Case
1
2
3
4
5
6
7
8
9
H
h2
h1
h2
h2
h1
h2
h1
h2
h1
S
s1
s1
s2
s2
s1
s2
s2
s2
s2
Case
1
2
3
4
5
6
7
8
9
V
v1
v2
v1
v2
v1
v1
v1
v1
v1
H
h1
h1
h2
h2
h2
h2
h1
h2
h1
S
s1
s1
s2
s2
s1
s2
s2
s2
s2
V
v1
v2
v1
v2
v1
v1
v1
v1
v1
They are identical except that in the table on the left we have
¡
¢
Case 1 = h2 s1 v1
Case 5 =
¡
h1
s1 v1
and in the table on the right we have
¡
Case 1 = h1 s1
Case 5 =
¡
h2
v1
s1 v1
¢
,
¢
¢
.
Clearly, P (di |GH ) will be the same for the these two di s since the value in
Corollary 6.6 does not depend on the order of the data. Similarly if, for example,
we flip around Case 2 and Case 3, we will not affect the result of the computation.
So, in general, for all di s which have the same data but in different order, we
need only compute P (di |GH ) once, and then multiply this value by the number
of such di s. As an example, consider again the di in the following table:
8.5. HIDDEN VARIABLE DAG MODELS
Case
1
2
3
4
5
6
7
8
9
H
h2
h1
h2
h2
h1
h2
h1
h2
h1
S
s1
s1
s2
s2
s1
s2
s2
s2
s2
489
V
v1
v2
v1
v2
v1
v1
v1
v1
v1
In this table, we have the following:
Value
¡
¢
¡ s1 v1 ¢
¡ s1 v2 ¢
¡ s2 v1 ¢
s2 v2
# of Cases
with this Value
# of Cases
with H Equal to h1
2
1
5
1
1
1
2
0
¡ ¢ ¡1¢ ¡5¢ ¡1¢
= 20 di s which have the same data as the one
So there are 21
1
2
0
above except in a different order. This means we need only compute P (di |GH )
for the di above, and multiply this result by 20.
Using this methodology, the following pseudocode shows the algorithm that
replaces the sum in Equality 8.30:
total = 0;
¡
for (k1 = 0; k1 <= M 1; k1 + +)
// M1 = # of ¡ s1
for (k2 = 0; k2 <= M2; k2 + +)
// M2 = # of ¡ s1
for (k3 = 0; k3 <= M 3; k3 + +)
// M3 = # of ¡ s2
for (k4 = 0; k4 <= M4; k4 + +) // M4 = # of s2
³ ´³ ´³ ´³ ´
1
M2
M3
M4
P (di |GH );
total = total + M
k1
k2
k3
k4
¢
v1 ¢ .
v2 ¢ .
v1 ¢ .
v2 .
// di is any data that has the following:
//
//
//
//
k1
k2
k3
k4
occurrences
occurrences
occurrences
occurrences
of
of
of
of
h1
h1
h1
h1
in
in
in
in
the
the
the
the
cases
cases
cases
cases
¡
¢
with ¡ s1 v1 ¢
with ¡ s1 v2 ¢
with ¡ s1 v2 ¢
with s2 v2 .
We have replaced exponential time complexity by quadratic time complexity.
Given the data shown at the beginning of this subsection, we compute 3 × 2 ×
6 × 2 = 72 values instead of 29 = 512 values.
490
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
Cooper [1995b] presents a general algorithm based on the method just presented that allows for more than two observables and allows variables (including
the hidden variable) to have arbitrary discrete space sizes. If we let
n
r
M
f
=
=
=
=
number of variables in network other than the single hidden variable
maximum space size of all variables
number of cases in the data
number of different instantiations of the variables,
he shows the number of values computed is in
´
³
f(r−1)
.
O M
f + r − 1)
(8.31)
In [Cooper and Herskovits, 1992], it is shown that the time complexity to compute any one term (i.e. P (di |GH )) is in
¢
¡
O M n2 r .
For example, in the case of the data presented at the beginning of this subsection,
n
r
M
f
=
=
=
=
2
2
9
4.
So the number of terms computed is in
O((M/4 + 2 − 1)4(2−1) ) = O(M 4 ),
which is quadratic in terms of M as we have already noted.
Although the time complexity shown in Expression 8.31 is polynomial, the
degree of the polynomial is large if either f or r is large.
More Then One Hidden Variable So far we have considered only one
hidden variable. Cooper [1995b] discusses extending the method just presented
to the case where we postulate more than one hidden variable.
Approximation Methods
The more efficient method for determining the Bayesian score of a hidden variable DAG model, which was developed in the previous subsection, is still computationally intensive. So approximation methods are sometimes more appropriate. Clearly, the Monte Carlo methods developed in Section 8.3.1 can be used to
approximate the Bayesian score of a hidden variable DAG model. Furthermore,
recall from Section 8.5.3 that Geiger et al [1996] show that, if we determine
the dimension d of the model as illustrated in that section, the scores discussed
8.6. LEARNING STRUCTURE: CONTINUOUS VARIABLES
491
in Section 8.3.2 approximate the Bayesian score in the case of hidden variable
DAG models. However, one score we developed must be modified. Recall, in
Section 8.3.2 we showed the Cheeseman-Stutz approximation is asymptotically
correct by showing
³
´
³
´
CS(d, G) ≡ ln (P (d0 |G)) − ln P (d0 |f̂ (G) , G) + ln P (d|f̂ (G) , G)
¸ ·
¸
·
d
d
0
0
= ln (P (d |G)) − BIC (d , G) + ln M + BIC (d, G) + ln M
2
2
= ln (P (d0 |G)) − BIC (d0 , G) + BIC (d, G) .
In this development we assumed that the dimension d0 of the model given data
d0 in the region of f̃ (G) is equal to the dimension d of the model given data d
in the region of f̃ (G) , and so the dimensions cancelled out. However, in the case
of hidden variable DAG models, d0 is the dimension of the model in which the
hidden variables are not hidden, while d is the dimension of the model in which
they are. As discussed in Section 8.5.3, we may have d0 > d. So in the case of
hidden variable DAG models, our Cheeseman-Stutz approximation must be as
follows:
³
³
´ d0
´ d
CS(d, G) ≡ ln (P (d0 |G))−ln P (d0 |f̂ (G) , G) + ln M+ln P (d|f̂ (G) , G) − ln M.
2
2
8.6
Learning Structure: Continuous Variables
An algorithm for doing inference in Gaussian Bayesian networks was developed
in Section 4.1. A method for learning parameters in such networks appears in
Section 7.2. Presently, we develop a method for learning structure assuming a
Gaussian Bayesian network.
Recall that given a set D = {X(1) , X(2) , . . . X(M) } of M random vectors such
(h)
that for every i all Xi have same space and a set d of values of the vectors in
D, in order to learn structure we need, for each value gp of GP , the conditional
density function
ρ(d|gp).
We first develop a method for obtaining the density function of D in the case
of a multivariate normal sample. Then we apply this result to determine the
density function of D given a DAG pattern gp.
8.6.1
The Density Function of D
Recall the definition of a multivariate normal sample (Definition 7.21) and Theorem 7.28. This theorem says for a multivariate normal sample D = {X(1) , X(2) ,
. . . X(M) } with prior parameter values β, α, µ, and v, and data d = {x(1) , x(2) ,
. . . x(M ) }, the updated parameter values β ∗ , α∗ , µ∗ , and v ∗ are as follows: If
492
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
we let
x=
PM
x(h)
h=1
and
M
s=
M ³
´³
´T
X
x(h) − x x(h) − x ,
h=1
then
β∗ = β + s +
vM
(x − µ)(x − µ)T
v+M
and
µ∗ =
vµ + Mx
v+M
and
α∗ = α + M,
and
(8.32)
v ∗ = v + M.
Now let dh = {x(1) , x(2) , . . . x(h) } where 1 ≤ h ≤ M. Denote the updated
parameters based on this subset of the data by β h , αh , µh , and vh . Owing to
the result in Theorem 7.28 if we let
Ph
h ³
´³
´T
(i)
X
i=1 x
x(i) − x x(i) − x ,
and
s=
x=
h
i=1
then
βh = β + s +
vh
(x − µ)(x − µ)T
v+h
and
αh = α + h,
(8.33)
and
vµ + hx
v+h
We have the following lemma:
and
µh =
vh = v + h.
Lemma 8.2 Given a multivariate normal sample D with prior parameters values β, α, µ, and v,
¶
µ
v
(x(h+1) − µ)(x(h+1) − µ)T .
βh+1 = β h +
v+1
Proof. The proof is left as an exercise.
We also need this lemma, which establishes a less traditional form of the t
density function.
Lemma 8.3 Suppose X is an n-dimensional random vector with density function
µ
¶
v(α − n + 1) −1
β
ρ(X) = t x; α − n + 1, µ,
.
(v + 1)
Then
ρ(X) =
µ
1
2π
¶n µ
2
v
v+1
¶n µ
2
c(n, α)
c(n, α + 1)
¶

α
2
|β|
,

α+1
2
|β̂|

8.6. LEARNING STRUCTURE: CONTINUOUS VARIABLES
where
β̂ = β +
and
"
µ
¶
v
(x − µ)(x − µ)T ,
v+1
αn/2 n(n−1)/4
c (n, α) = 2
493
π
µ
¶#−1
n
Y
α+1−i
Γ
.
2
i=1
(8.34)
Note that the function c (n, α) is the one obtained for the Wishart density function in Theorem 7.25.
Proof. The proof can be found in [Box and Tiao, 1973].
Now we can prove the theorem which gives the density function of D.
Theorem 8.8 Suppose we have a multivariate normal sample D = {X(1) , X(2) ,
. . . X(M ) }, where the X(h) s are n-dimensional, with prior parameter values
β, α, µ, and v, and data d = {x(1) , x(2) , . . . x(M) }. Then


α
µ ¶Mn µ
¶n µ
¶
2
2
2
1
v
c(n, α)
,
 |β|
ρ(d) =
α+M
2π
v+M
c(n, α + M)
|β ∗ | 2
where β ∗ is given by Equality 8.32 and c (n, α) is given by Equality 8.34. Note
that we do not condition on a DAG pattern because we now are talking only
about a multivariate normal sample.
Proof. Owing to the chain rule and Theorem 7.29, we have
X
ρ(d) =
ρ(x(1) , . . . x(M ) )
=
M
−1
Y
h=0
=
M
−1
Y
h=0
=
M
−1
Y
h=0
ρ(x(h+1) |x(1) , . . . x(h) )
µ
(h+1)
t x
vh (αh − n + 1)
(β h )−1
; αh − n + 1, µh ,
(vh + 1)
¶
µ
¶
(v + h) (α + h − n + 1)
(β h )−1 ,
t x(h+1) ; α + h − n + 1, µh ,
(v + h + 1)
where α0 , µ0 , v0 , and β 0 denote the initial values α, µ, v, and β. The last
equality is due to the right equality in Equality 8.33. Owing to Lemmas 8.2 and
??, we have


α+h
¶n µ
¶n µ
¶
M
−1 µ
Y
2
2
|
1 2
v+h
c(n, α + h)
|β
h
.

ρ(d) =
α+h+1
2π
v
+
h
+
1
c(n,
α
+
h
+
1)
h=0
|βh+1 | 2
Simplifying this product completes the proof.
494
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
We prove one more theorem, which will be needed in the next subsection. However, before proving it, we need to develop some notation. If V =
{X1 , . . . Xn }, W ⊆ V, and M is an n × n matrix,
M(W)
denotes the submatrix of M containing entries Mij such that Xi , Xj ∈ W. We
now have the following theorem:
Theorem 8.9 Suppose we have a multivariate normal sample D = {X(1) , X(2) ,
. . . X(M ) }, where the X(h) s are n-dimensional, with prior parameter values
β, α, µ, and v, and data d = {x(1) , x(2) , . . . x(M ) }. Let V = {X1 , . . . Xn }. For
W ⊆ V let
³¡
¢(W) ´−1
,
β W = β −1
β ∗W = βW + sW +
vM
(xW − µ)(xW − µ)T ,
v+M
where subscripting sW and xW with W means the quantities are computed for
the data restricted to the variables in W,
αW = α − n + lW
where lW is the number if elements in W, and
dW
denote the data d restricted to the variables in W. Then
ρ (dW ) =
µ
1
2π
¶ M lW µ
2
v
v+M
¶ lW µ
2


αW
¶
2
|
c(lW , αW )
|β
W
.

αW +M
c(lW , αW + M )
∗
|β W | 2
Proof. The proof can be found in [Heckerman and Geiger, 1995].
Note that subscripting with
the data items. For example, if
 (1)

 x1

d =  x(1)
2


(1)
x3
and W = {x1 , x3 }, then
dW =
(Ã
W restricts the components of the vectors, not
(1)
x1
x(1)
3
  (2) 
(2)

x1
x1

  (2)   (2) 
 ,  x2  ,  x2  ,


(2)
(2)
x3
x3
 
! Ã
,
(2)
x1
x(2)
3
! Ã
,
(2)
x1
x(2)
3
!)
.
8.6. LEARNING STRUCTURE: CONTINUOUS VARIABLES
8.6.2
495
The Density function of D Given a DAG pattern
First we obtain further results for augmented Bayesian networks in general.
Then we apply our results to learning structure.
Further Results for Augmented Bayesian Networks
We will use the notation ρ when referring to the joint distribution embedded
in an augmented Bayesian network because we will be applying these results to
continuous variables. We start with a lemma.
Lemma 8.4 Let an augmented Bayesian network (G, F(G) , ρ|G) be given where
G = (V, E) and V = {X1 , X2 , . . . Xn }. Then
ρ(x1 , x2 , . . . xn |f, G) =
n
Y
i=1
ρ(xi |pai , fi |G).
Proof. The proof is left as an exercise.
We have the following definitions:
Definition 8.8 Suppose we have a set of augmented Bayesian networks such
that
1. every DAG G in each of the networks contains the same set of variables
{X1 , X2 , . . . Xn };
2. for every i, if (G1 , F(G1 ) , ρ|G1 ) and (G2 , F(G2 ) , ρ|G2 ) are two networks in
(G )
(G )
the set such that Xi has the same parents in G1 and G2 , then Fi 1 = Fi 2 .
Then the set is called a class of augmented Bayesian networks.
Example 8.20 Any subset of all multinomial augmented Bayesian networks
such that the DAGs in the networks all contain the same variables is a class
of augmented Bayesian networks. Such a class is called a multinomial augmented Bayesian network class
Example 8.21 Any subset of all Gaussian augmented Bayesian networks such
that the DAGs in the networks all contain the same variables is a class of augmented Bayesian networks. Such a class is called a Gaussian augmented
Bayesian network class.
Definition 8.9 Likelihood Modularity holds for a class of augmented Bayesian networks if for every two networks (G1 , F(G1 ) , ρ|G1 ) and (G2 , F(G2 ) , ρ|G2 ) in
the class and every i, if Xi has the same parents in G1 and G2 , then for all
values pai and all values fi ,
ρ(xi |pai , fi , G1 ) = ρ(xi |pai , fi , G2 ).
(8.35)
where PAi is the parent set of Xi in the two DAGs, and Fi is the parameter set
associated with Xi in the two DAGs.
496
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
Example 8.22 It is left as an exercise to show likelihood modularity holds for
any Multinomial augmented Bayesian network class.
Example 8.23 It is left as an exercise to show likelihood modularity holds for
any Gaussian augmented Bayesian network class.
Definition 8.10 Parameter Modularity holds for a class of augmented Bayesian networks if for every two networks (G1 , F(G1 ) , ρ|G1 ) and (G2 , F(G2 ) , ρ|G2 ) in
the class and every i, if Xi has the same parents in G1 and G2 , then
ρ(fi |G1 ) = ρ(fi |G2 ),
(8.36)
where Fi is the parameter set associated with Xi in the two DAGs.
To illustrate parameter modularity suppose G1 is X1 → X2 ← X3 and G2
is X1 → X2 → X3 . Since X1 has the same parents in both (none), parameter
modularity would imply ρ(f1 |G1 ) = ρ(f1 |G2 ). Clearly, parameter modularity
does not hold for every Multinomial augmented Bayesian network class or for
every Gaussian augmented Bayesian network class. However, we do have the
following:
Example 8.24 It is left as an exercise to show parameter modularity holds for
any Multinomial augmented Bayesian network class satisfying the following:
1. The probability distributions in all the embedded Bayesian networks are
the same.
2. The specified density functions are all Dirichlet.
3. All the augmented networks have the same equivalent sample size.
Owing to the previous example, we see that parameter modularity holds for
the class of augmented Bayesian networks in a multinomial Bayesian network
structure learning schema (See Definition 8.1.). Like size equivalence (See Definition 8.7.) the following definition could be stated using only the Bayesian
network models contained in augmented Bayesian networks.
Definition 8.11 Distribution Equivalence holds for a class of augmented
Bayesian networks if for every two networks (G1 , F(G1 ) , ρ|G1 ) and (G2 , F(G2 ) ,
ρ|G2 ) such that G1 and G2 are Markov equivalent, for every value f (G1 ) of F(G1 )
there exists a value f (G2 ) of F(G2 ) such that
ρ(x1, x2 , . . . xn |f (G1 ) , G1 ) = ρ(x1, x2 , . . . xn |f (G2 ) , G2 ).
Example 8.25 It is left as an exercise to show distribution equivalence holds
for any Multinomial augmented Bayesian network class.
Example 8.26 Shachter and Kenley [1989] show distribution equivalence holds
for any Gaussian augmented Bayesian network class.
8.6. LEARNING STRUCTURE: CONTINUOUS VARIABLES
497
Example 8.27 Suppose we have a class of augmented Bayesian networks whose
set of variables consists of three or more binary variables, for each i F(G)
=
i
(Ai , Bi ) and
ρ(Xi = 1|pai , ai , bi , G) =
1 + exp(ai +
1
P
Xj ∈PAi bij xj )
.
We did not show the dependence of Ai and Bi on G for the sake of simplicity.
Distribution equivalence does not hold for this class. For example, if G1 contains
the edges X1 → X2 , X2 → X3 , and X1 → X3 , and G2 contains the edges
X1 → X2 , X3 → X2 , and X1 → X3 , then clearly G1 and G2 are Markov
equivalent. It is left as an exercise to show there are joint distributions which
can be obtained from a value of F(G1 ) but not from a value of F(G2 ) and vice
versa.
Suppose now we have a class of augmented Bayesian networks such that the
DAG G in each network contains the variables {X1 , X2 , . . . Xn }. Suppose further
we have a set D = {X(1) , X(2) , . . . X(M ) } of n-dimensional random vectors such
(h)
that each Xi has the same space as Xi , (See Definitions 6.11 and 7.4), and a
set d of values of the vectors in D (data). Then for any member (G, F(G) , ρ|G)
of the class we can update relative to d by considering D a Bayesian network
sample with parameter (G, F(G) ). We have the following definition, lemmas, and
theorem concerning updating:
Definition 8.12 Suppose we have a class of augmented Bayesian networks for
which distribution equivalence holds. Suppose further that for all data d, for
every two networks (G1 , F(G1 ) , ρ|G1 ) and (G2 , F(G2 ) , ρ|G2 ) in the class such that
G1 and G2 are Markov equivalent,
ρ(d|G1 ) = ρ(d|G2 ),
where ρ(d|G1 ) and ρ(d|G2 ) are the density functions of d when D is considered
a Bayesian network sample with parameters (G1 , F(G1 ) ) and (G2 , F(G2 ) ) respectively. Then we say Likelihood Equivalence holds for the class.
Example 8.28 Heckerman et al [1995] show likelihood equivalence holds for
any Multinomial augmented Bayesian network class satisfying the following:
1. The probability distributions in all the embedded Bayesian networks are
the same.
2. The specified density functions are all Dirichlet.
3. All the augmented networks have the same equivalent sample size.
The result in the previous example is the same result as in Lemma 7.4.
498
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
Lemma 8.5 (Posterior Parameter Modularity) Suppose we have a class
of augmented Bayesian networks for which likelihood modularity and parameter modularity hold. Then the parameters remain modular when we update
relative to and data d. That is, for every two networks (G1 , F(G1 ) , ρ|G1 ) and
(G2 , F(G2 ) , ρ|G2 ) in the class, for every i, if Xi has the same parents in G1 and
G2 , then
ρ(fi |d, G1 ) = ρ(fi |d, G2 ).
Proof. It is left as an exercise to use Equalities 6.6, 8.35, and 8.36 to obtain
the proof.
Lemma 8.6 Suppose we have a class of augmented Bayesian networks for
which likelihood modularity holds. Suppose further the DAG G in each network in the class contains the variables in V. For W ⊆ V let (GW , F(GW ) , ρ|GW )
be a member of the class for which GW is a complete DAG in which all variables
in W are ordered first. That is, no variable in V − W has an edge to a variable
in W. Then for any data d,
ρ(W|d, GW ) = ρ(W|dW , GW ).
As before, dW denotes the data d restricted to variables in W.
Proof. It is left as an exercise to use Equalities 6.6, and 8.35 to obtain the
proof.
Theorem 8.10 Suppose we have a class of augmented Bayesian networks for
which likelihood modularity, parameter modularity, and likelihood equivalence
hold. Suppose further (GC , F(GC ) , ρ|GC ) is a member of the class containing a
complete DAG GC . Then for any member (G, F(G) , ρ|G) of the class and data
d = {x(1) , x(2) , . . . x(M ) }
ρ(d|G) =
n ρ(d (G)
|GC )
Y
PA ∪{xi }
i
i=1
(G)
where PAi
ρ(dPA(G) |GC )
,
(8.37)
i
is the set of parents of Xi in G.
Proof. For the sake of notational simplicity we do not superscript parent sets
with a DAG in this proof. Let
d(h) = {x(1) , . . . x(h−1) }.
8.6. LEARNING STRUCTURE: CONTINUOUS VARIABLES
499
Owing to the chain rule, we have
ρ(d|G) =
M
Y
h=1
=
ρ(x(h) |d(h) , G)
M Z
Y
ρ(x(h) |f (G) , d(h) , G)ρ(f (G) |d(h) , G)df (G)
h=1
=
M Z
Y
ρ(x(h) |f (G) , G)ρ(f (G) |d(h) , G)df (G)
h=1
=
h=1
=
M
Y
h=1
=
"
M Z
Y
n
Y
i=1
i=1
(h) (G)
(G) (h)
ρ(x(h)
, G)
i |pai , fi , G)ρ(fi |d
Z "Y
n
M Y
n Z
Y
h=1 i=1
=
n Z
M Y
Y
h=1 i=1
#
(h)
(h) (G)
ρ(xi |pai , fi , G)
ρ(f (G) |d(h) , G)df (G)
#
(h)
(h)
(G)
, G)ρ(fi |d(h) , G)dfi
(h)
(h)
(G)
, Gi )ρ(fi |d(h) , Gi )df (G)
ρ(xi |pai , fi
ρ(xi |pai , fi
(G)
df (G)
(G)
(G)
(8.38)
where Gi is a complete DAG in which all variables in PAi are ordered first,
followed by Xi and then by the remaining variables. The third equality is because
the X(h) s are mutually independent conditional on f (G) , the fourth equality is
due to Lemma 8.4, the fifth equality is due to Theorem 6.9, and the seventh
equality follows from Equality 8.35 and Lemma 8.5.
We have after performing the integration in Equality 8.38 that
ρ(d|G) =
M Y
n
Y
(h) (h)
ρ(x(h)
, Gi )
i |pai , d
h=1 i=1
ρ(pai |d(h) , Gi )
h=1 i=1
=
=
M Y
n
(h)
(h)
Y
ρ(x , pa |d(h) , Gi )
i
n ρ(x(h) , pa(h) |d(h)
M Y
Y
i
i
PAi ∪{Xi } , Gi )
(h)
h=1 i=1
=
=
=
i
(h)
(h)
ρ(pai |dPAi , Gi )
M
n Y
(h)
(1)
(1)
(h−1)
Y
ρ(x(h)
, x(h−1)
, Gi )
i , pai |pai , xi , . . . pai
i
(1)
(h−1)
ρ(pa(h)
, Gi )
i=1 h=1
i |pai , . . . pai
n
(1)
(1)
(M
)
(M
)
Y ρ(pa , x , . . . pa , x |Gi )
i
i
i
i
(1)
(M )
ρ(pai , . . . pai |Gi )
i=1
n
Y
ρ(dPAi ∪{Xi } |Gi )
i=1
ρ(dPAi |Gi )
(8.39)
500
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
The third equality is due to Lemma 8.6, the fourth is obtained by writing the
(h)
elements of the sets d(h)
PAi and dPAi ∪{Xi } , and the fifth is due to the chain rule.
Due to likelihood equivalence we have for 1 ≤ i ≤ n
ρ(d|Gi ) = ρ(d|GC ).
So for any subset W ⊆ V we obtain, by summing over all values of the variables
in DV−W , that for 1 ≤ i ≤ n
ρ(dW |Gi ) = ρ(dW |GC ).
Applying this result to Equality 8.39 completes the proof.
Example 8.29 Owing to Examples 8.22, 8.24, and 8.28, likelihood modularity, parameter modularity, and likelihood equivalence hold for any Multinomial
augmented Bayesian network class satisfying the following:
1. The probability distributions in all the embedded Bayesian networks are
the same.
2. The specified density functions are all Dirichlet.
3. All the augmented networks have the same equivalent sample size.
Therefore, the result in the previous theorem holds for this class. Heckerman
and Geiger [1995] use the fact to obtain the result in Corollary 7.6, which we
obtained directly.
Learning Structure
First we show how to compute the density of d given a complete DAG pattern.
Then we show how to do it for an arbitrary DAG pattern.
A Complete DAG pattern Recall from Section 8.1.1 that a DAG pattern
event is the event that all and only the d-separations in the pattern are conditional independencies in the relative frequency distribution of the variables.
Recall further that a multivariate normal sample (Definition 7.21) makes no
assumptions of conditional independencies. So we assume that given the event
that no conditional independencies are present, we have a multivariate normal
sample. Since the complete DAG pattern event gpC is the event that no conditional independencies are present, owing to Theorem 8.8 we have


α
µ ¶Mn µ
¶n µ
¶
2
2
2
1
v
c(n, α)
,
 |β|
(8.40)
ρ(d|gpC ) =
α+M
2π
v+M
c(n, α + M)
|β ∗ | 2
where
"
αn/2 n(n−1)/4
c (n, α) = 2
π
µ
¶#−1
n
Y
α+1−i
Γ
.
2
i=1
8.6. LEARNING STRUCTURE: CONTINUOUS VARIABLES
F12 = 4/3
:1 = 0
X1
501
F22 = 4/3
:2 = 0
b21= 0
X2
Figure 8.15: A Gaussian Bayesian network
To compute the density function of the data given this pattern, we first use
Steps 1-5 in Section 7.2.3 to determine the values needed in Theorem 8.8. Then
we apply that theorem to obtain our density function. An example follows.
Example 8.30 Suppose we have variables X1 and X2 and these data d:
Case
1
2
3
4
X1
2
2
6
6
X2
4
8
4
8
To compute the density of the data given the complete DAG pattern, we first
use Steps 1-5 in Section 7.2.3 to determine the values needed in Theorem 8.8.
1. Construct a prior Gaussian Bayesian network (G, P ). Suppose it is the
network in Figure 8.15.
2. Convert the Gaussian Bayesian network to the corresponding multivariate
normal distribution using the algorithm in Section 7.2.3. We need only
use Equality 7.28 to do this. We obtain the density function N (x; µ, T−1 )
where
µ ¶
0
µ=
0
and
T−1 =
µ
4/3 0
0 4/3
¶
.
3. Assess prior values of α and.v. We assess
v =3
and
α = 2.
4. Compute the prior value of β. We have
β
=
=
v(α − n + 1) −1
T
(v + 1)
µ
¶ µ
3(2 − 2 + 1) 4/3 0
1
=
0 4/3
0
(3 + 1)
0
1
¶
.
502
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
Note that we could not model prior ignorance by taking v = 0, α = −1,
and β = 0 (as discussed following Theorem 7.28). The function c(n, α) is
not defined for α = −1, and the expression on the right in Equality 8.40
would be 0 if β were 0. The reason this happens is that the density function
of X is improper if we use the values that represent prior ignorance (See
Theorem 7.27.) So we have chosen the smallest ‘legal’ value of α, and a
prior Gaussian Bayesian network that yields the identity matrix for β.
5. Apply Theorem 7.28 to compute the updated value β ∗ . We obtain
µ 311 288 ¶
∗
7
7
β =
.
(8.41)
551
288
7
7
We now have
ρ(d|gpC )
=
µ
1
2π
¶ Mn µ
=
µ
1
2π
¶ 4×2
=
µ
1
2π
2
2
¶ 4×2
2
v
v+M
¶n µ
2
c(n, α)
c(n, α + M )
¶

α
|β| 2


α+M
|β ∗ | 2
´
³
¶¯ 2 ¯µ
µ
¶¯ 2+4
¶2 µ
¶ ¯µ
2
¯ 1 0 ¯ 2 ¯ 311 288 ¯− 2
3
c(2, 2)
7
7
¯ ¯
¯
¯
551
¯
3+4
c(2, 2 + 4) ¯ 0 1 ¯ ¯ 288
7
7
´
³
¶¯ 2 ¯µ
µ
¶¯− 2+4
¶2 µ
¶ ¯µ
288
¯
2
2
3
1/4π ¯¯ 1 0 ¯¯ 2 ¯¯ 311
¯
7
7
551
¯
3+4
1/96π ¯ 0 1 ¯ ¯ 288
7
7

= 1. 12 × 10−12 .
An Arbitrary DAG Pattern It is reasonable to construct a class of augmented Bayesian networks, which is used to learn structure, so that likelihood
modularity, parameter modularity, and likelihood equivalence all hold. Examples 8.22, 8.24, and 8.28 show they do indeed hold for the class of augmented Bayesian networks in a multinomial Bayesian network structure learning schema. Assuming these three conditions, Theorem 8.10 gives us the means
to calculate the conditional probability of d given an arbitrary DAG from the
conditional probability of d given the complete DAG pattern. However, as is the
case for learning parameters in a Gaussian Bayesian networks, which is discussed
in Section 7.2.3, we do not actually develop augmented Gaussian Bayesian networks when we learn structure. However, we use the result in Theorem 8.10 to
motivate our structure learning methodology. That is, we compute the probability given the complete DAG pattern as shown in the previous subsection,
and then we define the probability given an arbitrary DAG pattern by Equality
8.37. Since this equality involves DAGs and not DAG patterns, we need show
this equality yields the same result for Markov equivalent DAGs. The following
theorem implies this is the case:
8.6. LEARNING STRUCTURE: CONTINUOUS VARIABLES
503
Theorem 8.11 Suppose G1 and G2 are Markov equivalent DAGs containing n
nodes {X1 , X2 , . . . Xn }. Then for any function f
n
(G )
Y
f (PA 1 ∪ {xi })
i
(G1 )
i=1
f(PAi
)
=
n
(G )
Y
f (PA 2 ∪ {xi })
i
(G2 )
f(PAi
i=1
.
)
Proof. The proof can be found in [Geiger and Heckerman, 1994].
So if gp is not the complete DAG pattern, we proceed as follows: We first
computeρ(d|gpC ) as shown in the previous subsection. We then specify some
DAG G in the equivalence class represented by gp, and we define
ρ(d|gp) = ρ(d|G) ≡
n ρ(d (G)
|gpC )
Y
PA ∪{xi }
i
i=1
ρ(dPA(G) |gpC )
,
(8.42)
i
(G)
where PAi is the set of parents of Xi in G. The values in Equality 8.42 can be
computed using Theorem 8.9. Recall in that theorem we did not condition on
a DAG pattern. However, we have assumed the assumptions in that theorem
constitute the hypothesis of a complete DAG pattern. As was the case for
discrete variables, we call the expression in Equality 8.42 the Bayesian scoring
criterion scoreB , and is used to score both DAGs and DAG patterns.
Example 8.31 Suppose we have the data in Example 8.30. Let gpI be the
pattern in which there is no edge between X1 and X2 . Then Equality 8.42
becomes
¢ ¡
¢
¡
ρ(d|gpI ) = ρ d{X1 } |gpC ρ d{X2 } |gpC .
We will apply Theorem 8.9 to compute this value. First we need to compute
β {X1 } , β {X2 } , β ∗{X1 } , and β∗{X2 } . We have
So
¡ −1 ¢
β
=
µ
1 0
0 1
¶−1
=
µ
1 0
0 1
¶
.
¡ −1 ¢{X1 }
= (1)
β
and
β {X1 } =
³¡
¢{X1 } ´−1
β−1
= (1)−1 = (1) .
Given this result, it is not hard to see that β ∗{X1 } is simply (β ∗11 ). Similarly,
β ∗{X2 } is (β ∗22 ). So by looking at Equality ?? we ascertain that
µ
µ
¶
¶
311
551
β ∗{X1 } =
and
β ∗{X2 } =
.
7
7
Owing to Theorem 8.9 we now have
¢
¡
ρ d{X1 }
504
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
=
µ
1
2π
¶ Ml{X1 } µ
=
µ
1
2π
¶ 4×1 µ
=
µ
1
2π
2
2
¶ 4×1
2
v
v+M
¶ l{X1 } µ
2
c(l{X1 } , α{X1 } )
c(l{X1 } , α{X1 } + M)
¶


α{X1 }
2
|β
|
{X1 }



α{X1 } +M 
2
|β ∗{X1 } |
´
³
¯µ
¶
¶¯ 1+4
1 ¯ 311 ¯−
2
c(1, 1)
¯
¯
|1| 2 ¯
¯
c(1, 1 + 4)
7
³
´
¯µ
µ
¶ 1 Ã ¡√ √ ¢ !
¶¯ 1+4
1 ¯ 311 ¯−
2
2
1/ 2 π
3
¯
¡ √ √ ¢ |1| 2 ¯¯
3+4
7 ¯
1/ 3 2 π
3
3+4
¶1 µ
2
= 3. 78 × 10−6
¢
¡
ρ d{X2 }
=
µ
1
2π
¶ Ml{X2 } µ
=
µ
1
2π
¶ 4×1 µ
=
µ
1
2π
2
2
¶ 4×1
2
v
v+M
¶ l{X2 } µ
2
c(l{X2 } , α{X2 } )
c(l{X2 } , α{X2 } + M )
³
¶


α{X2 }
2
|β
|

{X2 }


α{X2 } +M 
2
|β∗{X2 } |
´
¶ 1 ¯µ
¶¯ 1+4
¯ 551 ¯− 2
c(1, 1)
¯
|1| 2 ¯¯
c(1, 1 + 4)
7 ¯
´
³
µ
¶ 1 à ¡√ √ ¢ ! 1 ¯µ
¶¯ 1+4
¯ 551 ¯− 2
2
1/ 2 π
3
¯
¡ √ √ ¢ |1| 2 ¯¯
3+4
7 ¯
1/ 3 2 π
3
3+4
¶1 µ
2
= 9. 05 × 10−7 .
We therefore have
¢ ¡
¢
¡
ρ(d|gpI ) = ρ d{X1 } ρ d{X2 }
¢¡
¢
¡
= 3. 78 × 10−6 9. 05 × 10−7
= 3. 42 × 10−12 .
Recall from Example 8.30 that
ρ(d|gpC ) = 1. 12 × 10−12 .
Note that gpI is about three times as likely as gpC . This is reasonable since the
data exhibits independence.
Example 8.32 Suppose we have two variables X1 and X2 variables and these
data d:
Case
1
2
3
4
X1
1
2
4
5
X2
1
2
4
5
8.7. LEARNING DYNAMIC BAYESIAN NETWORKS
505
To compute the density function of the data given the complete DAG pattern, we
first use Steps 1-5 in Section 7.2.3 to determine the values needed in Theorem
8.8. Suppose we use the same prior values as in Example 8.30. It is left as an
exercise to show
ρ(d|gpC ) = 4.73 × 10−8
ρ(d|gpI ) = 1. 92 × 10−10 .
Notice that gpC is much more likely. This is not surprising since the data exhibit
complete dependence.
Finally, we define the schema for learning structure in the case Gaussian
Bayesian networks. Given this schema definition, the definition of a Gaussian
Bayesian network structure learning space is analogous to Definition 8.2.
Definition 8.13 A Gaussian Bayesian network structure learning schema consists of the following:
1. a prior Gaussian Bayesian network (G, P );
2. values of α and v for a multivariate normal sample;
3. for each DAG pattern gp0 containing the variables in G a DAG G0 which
is any member of the equivalence class represented by gp0 .
[Monti, 1999] contains a Bayesian method for learning structure when the
network contains both discrete and continuous variables.
8.7
Learning Dynamic Bayesian Networks
Friedman et al [1998] developed methods for learning the structure of dynamic
Bayesian networks which mirror our results for Bayesian networks.
EXERCISES
Section 8.1
Exercise 8.1 Figure 9.6 shows the 11 DAG patterns containing three variables
X1 , X2 , and X3 . Create a Bayesian network structure learning schema for three
variables by assigning, for each variables Xi and each set of parents paij of Xi ,
a Dirichlet distribution of the variables in Fij such that for each k
aijk =
N
,
qi ri
506
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
where ri is the number of possible values of Xi , and qi is the number of different
instantiations of the parents of Xi . Do this for N = 1, 2, and 4. For each value
of N , create a schema with r1 = r2 = r3 = 2 and with r1 = 2, r2 = 3, r3 = 4.
Assign each DAG pattern a prior probability of 1/11.
Exercise 8.2 Suppose we create the joint probability distribution in a multinomial Bayesian network structure learning schema by assigning, for each variables Xi and each set of parents paij of Xi , a Dirichlet distribution of the variables in Fij such that for each k
aijk =
N
.
qi ri
Show that this results in the same probability being assigned to all combinations
of values of the Xi s.
Exercise 8.3 Suppose we have the Bayesian network learning schema in Exercise 8.1 with r1 = r2 = r3 = 2, and the following data:
Case
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
X1
1
1
1
1
1
1
1
1
1
2
2
2
1
1
2
2
X2
1
1
1
1
1
1
2
2
2
1
1
2
1
2
1
2
X3
1
1
1
1
1
1
1
1
1
1
1
1
2
2
2
2
Compute P ( gp|d) for each DAG pattern gp for each of the values of N . Is the
same DAG pattern most probable for all values of N ? Determine which, if any,
conditional independencies, hold for the data. Compare your determination to
the most probable DAG pattern(s).
Exercise 8.4 Suppose we have the Bayesian network learning schema in Exercise 8.1 with r1 = r2 = r3 = 2, and the following data:
EXERCISES
507
Case
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
X1
1
1
1
1
1
1
1
1
1
1
1
1
2
2
2
2
X2
1
1
1
1
1
1
1
1
1
2
2
2
1
1
1
2
X3
1
2
2
1
2
1
1
1
2
1
1
1
2
2
2
2
Compute P ( gp|d) for each DAG pattern gp and for each of the values of N . Is
the same DAG pattern most probable for all values of N ? Determine which, if
any, conditional independencies, hold for the data. Compare your determination
to the most probable DAG pattern(s).
Exercise 8.5 Show for recurrence 8.2 that f (2) = 3, f (3) = 25, f (5) = 29, 000,
and f (10) = 4.2 × 1018 .
Section 8.2
Exercise 8.6 Show for the problem instance discussed in Example 8.6 that
(9)
(9)
P (X1 = 2|X2 = 1, gp2 , d) = .41667.
(9)
(9)
Exercise 8.7 Using model averaging, compute P (X1 = 2|X2 = 1, d) given
the Bayesian network structure learning schema and data discussed a) in Example 8.3; and b) in Example 8.4.
Section 8.3
Exercise 8.8 Suppose we have a Markov chain with the following transition
matrix:


1/5 2/5 2/5
 1/7 4/7 2/7  .
3/8 1/8 1/2
Determine the stationary distribution for the chain.
508
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
¡
¢
Exercise 8.9 Suppose we have the distribution rT = 1/9 2/3 2/9 . Using the Metropolis-Hastings method, find a transition matrix for a Markov chain
that has this as its stationary distribution. Do it both with a matrix Q that is
symmetric and with one that is not.
Exercise 8.10 Implement the Candidate method, the Laplace approximation,
the BIC approximation, the MLED score, and the Cheeseman-Stutz approximation in the computer language of your choice, and compare their performance
using various data sets.
Section 8.5
Exercise 8.11 Use a probability distribution, which is faithful to the DAG in
Figure 8.7, to generate data containing values of N , F , C, and T , and determine
the Bayesian scores of models containing the DAGs in the equivalence classes
represented in Figure 8.6 and a model containing the DAG in Figure 8.7. Do
this for various sizes of the data set.
Section 8.6
Exercise 8.12 Prove Lemma 8.2.
Exercise 8.13 Prove Lemma 8.4.
Exercise 8.14 Show likelihood modularity holds for any Multinomial augmented
Bayesian network class.
Exercise 8.15 Show likelihood modularity holds for any Gaussian augmented
Bayesian network class.
Exercise 8.16 Show parameter modularity holds for any Multinomial augmented
Bayesian network class satisfying the following:
1. The probability distributions in all the embedded Bayesian networks are
the same.
2. The specified density functions are all Dirichlet.
3. All the augmented networks have the same equivalent sample size.
Exercise 8.17 Show distribution equivalence holds for any Multinomial augmented Bayesian network class.
Exercise 8.18 Prove Lemma 8.5.
EXERCISES
509
Exercise 8.19 Prove Lemma 8.6.
Exercise 8.20 Show that β ∗{X1 } = (β ∗11 ) in Example 8.31.
Exercise 8.21 Obtain the results in Example 8.32.
Exercise 8.22 Suppose we have variables X1 and X2 and these data d:
Case
1
2
3
4
X1
4
4
2
2
X2
1
3
1
5
Making the assumptions in a Gaussian Bayesian network and assuming the
priors in Example 8.30, determine the density of the data given the complete
DAG pattern and the DAG pattern with an edge between the variables.
Exercise 8.23 Given the data in Example 8.22 and the assumptions in a Gaussian
Bayesian network, ascertain priors different than those in Example 8.30, and
determine the density of the data given the complete DAG pattern and the DAG
pattern with an edge between the variables.
Exercise 8.24 Suppose we have variables X1 , X2 , and X3 , and these data d:
Case
1
2
3
4
X1
1
1
2
2
X2
5
6
7
8
X3
3
4
3
4
Making the assumptions in a Gaussian Bayesian network, ascertain priors
which come as close as possible to modeling prior ignorance, and determine the
density of the data given all possible DAG patterns.
510
CHAPTER 8. BAYESIAN STRUCTURE LEARNING
Chapter 9
Approximate Bayesian
Structure Learning
As discussed in Section 8.1.4, when the number of variables is not small, to
find the maximizing DAG patterns by exhaustively considering all DAG patterns is computationally unfeasible. Furthermore, in this case it is not possible
to average over all DAG patterns either. Section 9.1 discusses approximation
algorithms for finding the most probable structure, while Section 9.2 presents
algorithms for doing approximate model averaging.
9.1
Approximate Model Selection
Recall the problem of finding the most probable structure is called model selection, and that the purpose of model selection is to learn a DAG pattern
that can be used for inference and decision making. Given that we assign equal
prior probabilities to all DAG patterns, this amounts to finding the values of gp
that maximize P (d|gp), which we call the Bayesian score scoreB (d, gp). There
could be more than one such pattern. However, to simplify our discussion, we
assume there is a unique one. As mentioned above, to find this DAG pattern by
exhaustively considering all DAG patterns is computationally unfeasible when
the number of variables is not small.
One way to handle a problem like this is to develop a heuristic search algorithm. Heuristic search algorithms are algorithms that search for a solution
which is not guaranteed to be optimal (in this case the value of gp that maximizes scoreB (d, gp)), but rather they often find solutions that are reasonably
close to optimal. A search algorithm requires a search space which contains
all candidate solutions, and a set of operations that transforms one candidate
solution to another. In the case of learning Bayesian networks, perhaps the
simplest search space consists of all DAGs containing the n variables. We first
present a heuristic search algorithm whose search space is the set of such DAGs
and which requires a prior ordering of the variables. Then we present an algo511
512 CHAPTER 9. APPROXIMATE BAYESIAN STRUCTURE LEARNING
rithm which does not have the restriction of a prior ordering, but whose search
space is still a set of DAGs. Since we are really concerned with learning DAG
patterns, we then present an algorithm whose search space is all DAG patterns
containing the n variables. Finally, we present an approximate model selection
algorithm specifically for the cases of missing data and hidden variables.
Before proceeding we review the formulas we’ve developed for scoreB (d, gp),
and we develop an expression which shows the score as a product of local scores.
In the case of discrete variables Equality 8.1 gives scoreB (d, gp). Recall that
this equality says
(G)
scoreB (d, gp) = scoreB (d, G) =
n qY
i
Y
i=1 j=1
(G)
Γ(Nij )
(G)
(G)
(G)
(G)
ri
Y
Γ(aijk + sijk )
(G)
Γ(Nij + Mij ) k=1
,
Γ(aijk )
(9.1)
P (G)
(G)
(G)
(G)
(G)
where aij1 , aij2 , . . . aijri , and Nij = k aijk are their values in the multinomial augmented Bayesian network (G, F(G) , ρ|G) corresponding to gp.
In the case of continuous variables Equality 8.42 gives ρ(d|gp). Recall that
this equality says
scoreB (d, gp) = scoreB (d, G) =
n ρ(d (G)
|gpC )
Y
PA ∪{xi }
i
i=1
ρ(dPA(G) |gpC )
,
(9.2)
i
where G is any DAG in the equivalence class represented by gp, PA(G)
is the set
i
of parents of Xi in G, and gpC is the complete DAG pattern.
The score of a DAG or DAG pattern is the product of factors which locally
score each node paired with the node’s parents in the DAG. We show this next.
In the discrete case, for a given node Xi and parent set PA, set
q(PA)
scoreB (d, Xi , PA) =
Y
j=1
P (PA)
(PA)
(PA)
ri
Y
Γ( k aijk )
Γ(aijk + sijk )
,
P (PA) P (PA)
(PA)
Γ( k aijk + k sijk ) k=1
Γ(aijk )
(9.3)
(PA)
where q (PA) is the number of different instantiations of the variables in PA, sijk
is the number of cases in which Xi = k and in which the variables in PA are
(PA)
in their jth instantiation, etc. Note that aijk depends only on i, j, and k and
a parent set PA; it does not depend on a DAG. This follows from parameter
modularity (See Definition 8.10.). Similarly, in the continuous case set
scoreB (d, Xi , PA) =
ρ(dPA∪{xi } |gpC )
.
ρ(dPA |gpC )
(9.4)
In both the discrete and continuous cases, we then have
scoreB (d, gp) = scoreB (d, G) =
n
Y
i=1
(G)
where PAi
is the set of parents of Xi in G.
(G)
scoreB (d, Xi , PAi ),
9.1. APPROXIMATE MODEL SELECTION
9.1.1
513
Algorithms that Search over DAGs
We present two algorithms in which the search space consists of DAGs.
The K2 Algorithm
First we give the algorithm. Then we show an example in which it was applied.
The Algorithm Cooper and Herskovits [1992] developed a greedy search algorithm that searches for a DAG G0 that approximates maximizing scoreB (d, G).
That is, the search space is the set of all DAGs containing the n variables. Towards approximately maximizing the scoreB (d, G), for each i they locally find
a value PAi that approximately maximizes scoreB (d, Xi , PA). The single operation in this search algorithm is the addition of a parent to a node. The algorithm
proceeds as follows: We assume an ordering of the nodes such that if Xi precedes
Xj in the order, an arc from Xj to Xi is not allowed. Let P red(Xi ) be the set of
nodes that precede Xi in the ordering, We initially set the parents PAi of Xi to
empty and compute scoreB (d, Xi , PA). Next we visit the nodes in sequence according to the ordering. When we visit Xi , we determine the node in P red(Xi )
which most increases scoreB (d, Xi , PA). We ‘greedily’ add this node to PAi . We
continue doing this until the addition of no node increases scoreB (d, Xi , PA).
Pseudocode for this algorithm follows. The algorithm is called K2 because it
evolved from a system name Kutató [Herskovits and Cooper, 1990].
Algorithm 9.1 K2
Problem: Find a DAG G0 that approximates maximizing scoreB (d, G).
Inputs: A Bayesian network structure learning schema BL containing n variables, an upper bound u on the number of parents a node may have, data
d.
Outputs: n sets of parent nodes PAi , where 1 ≤ i ≤ n, in a DAG that
approximates maximizing scoreB (d, G).
void K2 (Bayes_net_struct_learn_schema BL, int u,
data d, for 1 ≤ i ≤ n parent_set& PAi )
{
for (i = 1; i <= n; i + +) {
// n is the number of nodes.
PAi = ∅;
Pold = scoreB (d, Xi , PAi );
f indmore = true;
while (f indmore && |PAi | < u) { // | | returns the size of a set.
514 CHAPTER 9. APPROXIMATE BAYESIAN STRUCTURE LEARNING
Z = node in P red(Xi ) − PAi that maximizes scoreB (d, Xi , PAi ∪ {Z});
Pnew = scoreB (d, Xi , PAi ∪ {Z}) ;
if (Pnew > Pold ) {
Pold = Pnew ;
PAi = PAi ∪ {Z};
}
else
findmore = false;
}
}
}
Next we analyze the algorithm:
Analysis of Algorithm 9.1 (K2)
Let
n
r
u
M
L
= number of variables (nodes)
= maximum number of values for any one variable
= upper bound on number of parents a node may have
= number of cases in the data
= maximum(Nij )
We assume the values of the gamma functions in Equality 9.3 are
first computed and stored in an array. There are no such values
greater than Γ(L+M ) in Equality 9.3 because Mij can have no value
greater than M . The time complexity of computing and storing the
factorials of the integers from 1 to L + M − 1 is in
O(L + M − 1).
In each iteration of the while loop, the determination of the value
of Z requires at most n − 1 computations of g because each node
has at most n − 1 predecessors. [Cooper and Herskovits, 1992] show
the time complexity to compute g once is in O(Mur). Therefore,
the time complexity to compute the value of Z once is in O(nMur).
The other statements in the while loop require constant time. There
are always at most u iterations of the while loop and n iterations
of the for loop. Therefore, the time complexity of the for loop is in
O(n2 M u2 r). Since u ≤ n, this time complexity of the for loop is in
O(n4 Mr).
9.1. APPROXIMATE MODEL SELECTION
515
You might wonder where we could obtain the ordering required by Algorithm
9.1. Such an ordering could be possibly be obtained from domain knowledge
such as a time ordering of the variables. For example, we might know that in
patients, smoking precedes bronchitis and lung cancer, and that each of these
conditions precedes fatigue and a positive chest X-ray.
An Example: The ALARM Network Cooper and Herskovits [1992] tested
the K2 algorithm in the following way. They randomly generated 10,000 cases
from the ALARM Bayesian network, used the K2 algorithm to learn a DAG
from the generated data, and then they compared the resultant DAG to the one
in the ALARM Bayesian network. We discuss their results next.
The ALARM Bayesian network ([Beinlich et al, 1989]) is an expert system
for identifying anesthesia problems in the operating room. It is shown in Figure
9.1. For the sake of brevity, we identify only a few values of the variables.
You are referred to the original source for the entire network. We identify the
following:
Variable
X8
X20
X27
X29
Value
1
1
1
1
Outcome Mapped to this Value
EKG shows the patient has an increased heart rate.
Patient is receiving insufficient anesthesia or analgesia.
Patient has an increased level of adrenaline.
Patient has an increased heart rate.
There are 37 nodes and 46 edges in the network. Of these 37 nodes, 8 are diagnostic problems, 16 are findings, and 13 are intermediate variables connecting
diagnostic problems to findings. Each node has from two to four possible values. Beinlich constructed the network from his personal domain knowledge of
the relationships among the variables.
Cooper and Herskovits [1992] generated 10,000 cases from the ALARM network using a Monte Carlo technique developed in [Henrion, 1988]. Their Monte
Carlo technique is an unbiased generator of cases in that the probability of a
particular case being generated is equal to the probability of that case according
to the Bayesian network. They supplied the K2 algorithm with the 10,000 cases
and an ordering of the nodes that is consistent with the partial ordering of the
nodes according to the DAG in the ALARM Bayesian network. For example,
X21 must appear before X10 in the ordering because there is an edge from X21
to X10 . However, they need not be consecutive. Indeed, X21 could be first and
X10 could be last. They manually generated the ordering by adding a node to
the list only when the node’s parents were already in the list. In their decision
as to where to place the nodes in the list, no regard was given to the meaning
of the nodes. Their resultant ordering is as follows (We only list the numbers
of the nodes.):
12, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 28, 30, 31, 37, 1,
2, 3, 4, 10, 36, 13, 35, 15, 34, 32, 33, 11, 14, 27, 29, 6, 7, 8, 9, 5
516 CHAPTER 9. APPROXIMATE BAYESIAN STRUCTURE LEARNING
X6
X17
X25
X18
X26
X1
X2
X3
X5
X10
X21
X22
X19
X20
X31
X15
X4
X27
X11
X28
X29
X7
X8
X32
X34
X13
X23
X35
X12
X9
X16
X36
X37
X24
X33
X14
X30
Figure 9.1: The DAG in the ALARM Bayesian network.
From the this ordering and the 10,000 cases, the K2 Algorithm constructed
a DAG identical to the one in the ALARM network, except that the edge from
X12 to X32 was missing, and an edge from X15 to X34 was added. Subsequent
analysis showed that the edge X12 to X32 was not strong supported by the
10,000 cases. [Cooper and Herskovits, 1992] presented the K2 algorithm with
the 100, 200, 500, 1000, 2000, and 3000 cases in the 10,000 case data file. They
found that the algorithm learned the same DAG from the first 3000 cases as it
learned from the entire file.
An Algorithm Without a Prior Ordering
First we present the algorithm; then we discuss improving it.
The Algorithm We present a straightforward greedy search that does not
require a time ordering. The search space is again the set of all DAGs containing
the n variables, and the set DAGOPS of operations is as follows:
1. If two nodes are not adjacent, add an edge between them in either direction.
2. If two nodes are adjacent, remove the edge between them.
3. If two nodes are adjacent, reverse the edge between them.
9.1. APPROXIMATE MODEL SELECTION
517
All operations are subject to the constraint that the resultant graph does not
contain a cycle. The set of all DAGs which can be obtained from G by applying
one of the operations is called Nbhd(G). If G0 ∈ Nbhd(G), we say G0 is in
the neighborhood of G . Clearly, this set of operations is complete for the
search space. That is, for any two DAGs G and G0 there exists a sequence of
operations that transforms G to G0 . The reverse edge operation is not needed
for the operations to be complete, but it increases the connectivity of the space
without adding too much complexity, which typically leads to better search.
Furthermore, when using a greedy search algorithm, including edge reversals
seems to often lead to a better local maximum.
The algorithm proceeds as follows: We start with a DAG with no edges. At
each step of the search, of all those DAGs in the neighborhood of our current
DAG, we ‘greedily’ choose the one that maximizes scoreB (d, G). We halt when
no operation increases this score. Note that in each step, if an edge to Xi is
added or deleted, we need only re-evaluate scoreB (d, Xi , PAi ) (See Equalities
9.3 and 9.4.). If an edge between Xi and Xj is reversed, we need only reevaluate scoreB (d,Xi , PAi ) and scoreB (d, Xj , PAj ). When a model searching
algorithm need only locally re-compute a few scores to determine the score for
the next model under consideration, we say the algorithm has local scoring
updating. A model with local scoring updating is considerably more efficient
than one without it. Note that Algorithm 9.1 also has local scoring updating.
The algorithm follows:
Algorithm 9.2 DAG Search
Problem: Find a DAG G that approximates maximizing scoreB (d, G).
Inputs: A Bayesian network structure learning schema BL containing n variables, data d.
Outputs: A set of edges E in a DAG that approximates maximizing scoreB (d, G).
void DAG_search (Bayes_net_struct_learn_schema BL,
data d,
set_of_edges& E)
{
E = ∅; G = (V, E);
do
if (any DAG in the neighborhood of our current DAG
increases scoreB (d, G))
modify E according to the one that increases scoreB (d, G) the most;
while (some operation increases scoreB (d, G));
}
518 CHAPTER 9. APPROXIMATE BAYESIAN STRUCTURE LEARNING
An Improvement to the Algorithm A problem with a greedy search algorithm is that it could halt at a candidate solution that locally maximizes the
objective function rather than globally maximizes it (See [Xiang et al, 1996].)
One way for dealing with this problem is iterated hill-climbing. In iterated hillclimbing, local search is done until a local maximum is obtained. Then, the
current structure is randomly perturbed, and the process is repeated. Finally,
the maximum over local maxima is used. Other methods for attempting to
avoid local maxima include simulated annealing [Metropolis et al, 1953], bestfirst search [Korf, 1993], and Gibb’s sampling (discussed in Section 8.3.1).
9.1.2
Algorithms that Search over DAG Patterns
Although Algorithms 9.1 and 9.2 find a DAG G rather than a DAG pattern, we
can use them to find a DAG pattern by determining the DAG pattern gp representing the Markov equivalence class to which G belongs. Since scoreB (d, gp) =
scoreB (d, G)), we have approximated maximizing scoreB (d, gp). However, as
discussed in [Anderson et al, 1995], there are a number of potential problems in
searching for a DAG instead of a DAG pattern. Briefly, we discuss two of the
problems. The first is efficiency. By searching over DAGs, the algorithm can
waste time encountering and rescoring DAGs in the same Markov equivalence
class. A second problem has to do with priors. If we search over DAGs, we
are implicitly assigning equal priors to all DAGs, which means DAG patterns
containing more DAGs will have higher prior probability. For example, if there
are n nodes, the complete DAG pattern (representing no conditional independencies) contains n! DAGs, whereas the pattern with no edges (representing
all variables are mutually independent) contains just one DAG. On the other
hand, recall from Section 8.1.4 that Gillispie and Pearlman [2001] show that an
asymptotic ratio of the number of DAGs to DAG patterns equal to about 3.7
is reached when the number of nodes is only 10. Therefore, on the average the
number of DAGs in a given equivalence class is small and perhaps our concern
about searching over DAGs is not necessary. Contrariwise, in simulations performed by Chickering [2001] the average number of DAGs, in the equivalence
classes over which his algorithm searched, were always greater than 8.5 and in
one case was 9.7 × 1019 .
When performing model selection, assigning equal priors to DAGs is not
necessarily a serious problem as we will finally select a high-scoring DAG which
corresponds to a high-scoring DAG pattern. However, as discussed in Section
9.2.2 (which follows), it can be a serious problem in the case of model averaging.
After presenting an algorithm that searches over DAG patterns, we improve
the algorithm.
A Basic Algorithm
Chickering [1996b] developed an algorithm which searches over DAG patterns.
After presenting the algorithm, we compare its results with those of Algorithm
9.2.
9.1. APPROXIMATE MODEL SELECTION
X1
X2
519
X1
X3
X2
X1
X3
X2
X3
X4
X4
X4
(a)
(b)
(c)
X1
X2
X1
X3
X2
X3
X4
X4
(d)
(e)
Figure 9.2: The DAG in (b) is a consistent extension of the PDAG in (a), while
the DAG in (c) is not. Each of the PDAGs in (d) and (e) does not admit a
consistent extension.
First we need some definitions. A PDAG (partially directed DAG) gp is
a graph containing both directed and undirected edges. A DAG G is called a
consistent extension of PDAG gp if G has the same nodes and links (edges
without regard to direction) as gp, all edges directed in gp are directed in G,
and G does not contain any uncoupled head-to-head meetings that are not in
gp. PDAG gp admits a consistent extension if there is at least one consistent
extension of gp. Notice that a DAG pattern (defined at the end of Section 2.2)
is a PDAG, and any DAG in the Markov equivalence class it represents is a
consistent extension of it.
Example 9.1 The DAG in Figure 9.2 (b) is a consistent extension of the
PDAG in Figure 9.2 (a), while the DAG in Figure 9.2 (c) is not because it
contains the uncoupled head-to-head meeting X2 → X1 ← X3 . The PDAG in
Figure 9.2 (d) does not admit a consistent extension because it is not possible
to direct the edges so as to have no new uncoupled head-to-head meetings, and
the PDAG in Figure 9.2 (e) does not admit a consistent extension because it
520 CHAPTER 9. APPROXIMATE BAYESIAN STRUCTURE LEARNING
is not possible to direct the edges so as to have no new uncoupled head-to-head
meetings without creating cycle.
Now we can describe the algorithm. The search space is the set of all DAG
patterns containing the n variables. We start with the following set of operations, which we call PDAGOPS:
1. If two nodes are not adjacent, add an edge between them that is undirected
or is directed in either direction.
2. If there is an undirected edge, delete it.
3. If there is a directed edge, delete it or reverse.
4. If there is an uncoupled undirected meeting Xi − Xj − Xk , add the uncoupled head-to-head meeting Xi → Xj ← Xk .
Anderson el al [1995] show this set of operators is complete for the search space.
That is, for any two DAG patterns gp1 and gp2 there exists a sequence of
operations that transforms gp1 to gp2 . As is the case for DAGs, the reverse
edge operation is not needed for the operations to be complete.
Note that the operations in PDAGOPS do not necessarily yield a PDAG that
is a DAG pattern. Next we show how to obtain a DAG pattern after applying
one of these operations. Assume we have the following two routines:
void f ind_DAG (PDAG gp, DAG& G, bool& switch)
void f ind_DAG_pattern(DAG G, DAG—pattern& gp)
If gp admits a consistent extension, routine f ind_DAG returns a consistent
extension G and sets switch to true; otherwise it sets switch to false. Routine
f ind_DAG_pattern returns the DAG pattern which represents the Markov
equivalence class to which G belongs. A θ(nm) algorithm, where n is the
number of nodes and m is the number of edges, for f ind_DAG appears in
[Dor and Tarsi, 1992]. To obtain a θ(m) algorithm for find_DAG_pattern,
where m is the number of edges in G, we first change all the edges in G, that
are not involved in an uncoupled head-to-head meeting, to undirected edges,
and then we apply the while loop in Algorithm 10.1 in Section 10.1.
We use the preceding two routines to perform an operation that transforms
DAG pattern gp to DAG pattern gp00 as follows:
Apply one of the operations in PDAGOPS to gp to obtain gp0 ;
f ind_DAG (gp0 , G, switch);
if (switch)
f ind_DAG_pattern( G, gp00 );
Now if an edge insertion results in a DAG pattern, but the edge in the DAG
pattern does not have the same direction as it had when inserted, we do not
9.1. APPROXIMATE MODEL SELECTION
X1
X2
X3
521
X1
X4
X2
X3
X5
X5
(a)
(b)
X1
X2
X3
X1
X4
X4
X2
X3
X5
X5
(c)
(d)
X4
Figure 9.3: An initial DAG pattern gp is in (a), the result gp0 of deleting the
edge X2 → X3 is in (b), the output G of find_DAG is in (c), and the output
gp00 of f ind_DAG_pattern is in (d).
want to allow the operation. That is, if the inserted edge is undirected it must
be undirected in the resultant DAG pattern, and if it is directed it must be
directed in the resultant DAG pattern. With this restriction, a DAG pattern
cannot directly yield another DAG pattern in two different ways. The set of all
allowable DAG patterns which can be obtained from gp is called Nbhd(gp).
Figure 9.3 shows an applications of one operation. That is, Figure 9.3 (a)
shows an initial DAG pattern gp, Figure 9.3 (b) shows the result gp0 of deleting
the edge X2 → X3 , Figure 9.3 (c) shows the output G of find_DAG, and Figure
9.3 (d) shows the output gp00 of f ind_DAG_pattern. Owing to our restriction,
we would not allow the edge X1 → X5 be added to the DAG pattern in Figure
9.3 (d) because it would end up being undirected in the resultant DAG pattern.
However, we would allow the edge X1 → X5 to be added to the DAG pattern
in Figure 9.3 (a) because it would end up being directed in the resultant DAG
522 CHAPTER 9. APPROXIMATE BAYESIAN STRUCTURE LEARNING
pattern.
Next we show the algorithm.
Algorithm 9.3 DAG Pattern Search
Problem: Find a DAG pattern gp that approximates maximizing scoreB (d, gp).
Inputs: A Bayesian network structure learning schema BL containing n variables, data d.
Outputs: A set of edges E in a DAG pattern that approximates maximizing
scoreB (d, gp).
void DAG_pattern_search (Bayes_net_struct_learn_schema BL,
data d,
set_of_edges& E)
{
E = ∅; gp = (V, E);
do
if (any DAG pattern in the neighborhood of our current DAG pattern
increases scoreB (d, gp))
modify E according to the one that increases scoreB (d, gp) the most;
while (some operation increases score(gp));
}
Note that the output G of find_DAG can be used to compute the conditional probability of the output gp00 of f ind_DAG_pattern. Note further that,
when we apply an operator to obtain a new element of the search space, we can
get a DAG pattern which is globally quite different from the previous pattern.
The example in Figure 9.3 illustrates this. That is, the algorithm does not
have local scoring updating. Recall that Algorithm 9.2 does have local scoring
updating.
Chickering [1996b] compared Algorithms 9.2 and 9.3. The performance results appear in Tables 9.1 and 9.2. We discuss these results next. A gold
standard Bayesian network containing binary variables, whose DAG contained
the number of nodes indicated, was used to generate data items. Using a structure learning schema in which the equivalent sample size was 8, Algorithms 9.2
and 9.3 were then used to learn a DAG and a DAG pattern respectively. The
DAG pattern, representing the Markov equivalence class to which the gold standard DAG belongs, was then compared to the DAG pattern, representing the
Markov equivalence class to which the DAG learned by Algorithm 9.2 belongs,
and to the DAG pattern learned by Algorithm 9.3. For each pair of nodes, if
the edges between these two nodes were different in the two DAG patterns, one
was added to the structural difference. This was done for 9 random data sets,
and the averages appear in Table 9.1 in the columns labeled Alg 9.2 Struct. and
9.1. APPROXIMATE MODEL SELECTION
# of
Nodes
5
10
15
20
25
Alg 9.3
Score
−1326.64
−2745.55
−3665.29
−5372.94
−6786.83
Alg 9.2
Score
−1327.06
−2764.95
−3677.17
−5408.67
−6860.24
Score
Diff.
0.42
18.50
11.88
35.73
73.41
523
Alg 9.3
Struct.
0.78
4.44
17.67
25.11
32.67
Alg 9.2
Struct.
1.44
10.56
21.89
30.78
47.11
Struct.
Diff.
0.66
6.12
4.22
5.67
14.44
Table 9.1: A comparison of performance of Algorithms 8.2 and 8.3. Each entry
is the average over 9 random data sets consisting of 500 items each.
Alg 9.3
Score
−101004
Alg 9.2
Score
−101255
Score
Diff.
251
Alg 9.3
Struct.
36.3
Alg 9.2
Struct.
51.5
Struct.
Diff.
15.2
Table 9.2: A comparison of performance of Algorithms 8.2 and 8.3 for the
ALARM Network. Each entry is the average over 10 random data bases consisting of 10,000 items each.
Alg 9.3 Struct. The structural difference appearing in the last column of the table is the structural difference for Algorithm 9.3 minus the structural difference
for Algorithm 9.2. So a positive structural difference indicates Algorithm 9.2
learned structures closer to that of the gold standards. The scores appearing
in the table are the averages of the logs of the probabilities of the data given
the structures learned. The score difference is the score of Alg. 9.2 minus the
score of Alg. 9.3. So a positive score difference indicates Algorithm 9.2 learned
structures that make the data more probable. We see that in every case Algorithm 9.3 outperformed Algorithm 9.2 according to both criteria. Table 9.2
contains the same information when the ALARM network was used to generate
the data. Again Algorithm 9.3 outperformed Algorithm 9.2. Note however that
Algorithm 9.3 performed somewhat worse than the K2 Algorithm (Algorithm
9.1) relative to the ALARM network. That is, Algorithm 9.3 had a structure
difference of 36.3 compared to 2 (See Section 9.1.1.) for the K2 Algorithm. This
is not surprising since the K2 Algorithm has the benefit of a prior ordering.
# of
Nodes
5
10
15
20
25
Alg 9.3
Time
1
18.11
70.44
184.67
487.33
Alg 9.2
Time
0
1.67
6.22
11.78
22.56
Time
Ratio
−
10.84
11.32
15.68
21.60
Table 9.3: A comparison of time usage of Algorithms 8.2 and 8.3. Each entry
is the average over 9 random data bases consisting of 500 items each.
524 CHAPTER 9. APPROXIMATE BAYESIAN STRUCTURE LEARNING
Table 9.3 shows results of Chickering’s [1996b] time performance comparison
of the two algorithms. Recall that both algorithms terminate when no operation
increases the score. The time shown in the table is the time until that happens.
It is not surprising that Algorithm 9.3 takes significantly longer time. First Algorithm 9.2 need only consider n(n−1) node pairs in each step, while Algorithm
9.3 needs consider n(n − 1) node pairs plus 2e(n − 2) uncoupled head-to-head
meetings in each step. Second, Algorithm 9.3 requires additional overhead by
calling f ind_DAG and f ind_DAG_pattern. Third, Algorithm 9.3 does not
have local scoring updating, whereas Algorithm 9.2 does. This last problem is
removed with the improvement in the next subsection.
Spirtes and Meek [1995] also develop a heuristic algorithm that searches for
a DAG pattern.
An Improvement to the Algorithm
A problem with Algorithm 9.3 is it does not have local scoring updating. In 2001
Chickering improved the algorithm by determining a way to update the score
locally. Table 9.4 shows the local change that is needed after each operation is
performed. You should consult the original source for a proof that the changes
indicated in Table 9.4 are correct. We only explain the notation in the table. In
a PDAG, nodes X and Y are called neighbors if there is an undirected edge
between them. In Table 9.4, PAX denotes the set of parents of X, NX denotes
the set of neighbors of X, NX,Y denotes the set of common neighbors of X and
Y , and ΩX,Y denotes PAX ∩NY . Furthermore, for any set M, M+X is shorthand
for M∪{X} and M−X is shorthand for M−{X}. Chickering [2001] also develops
necessary and sufficient conditions for each operation to be valid (That is, the
operation yields a DAG pattern). The score in Table 9.4 is the expression in
either Equalities 9.3 or 9.4.
When we use Table 9.4 to update the score the of new DAG pattern in Algorithm 9.3, we call this the improved Algorithm 9.3. Chickering [2001] compared
the improved Algorithm 9.3 to Algorithm 9.2 using six real-life data sets. Table
9.5 shows the results of the time comparisons. We see that in four of six cases
the improved Algorithm 9.3 terminated more rapidly, and in two of those cases
it terminated almost three times as fast. Chickering [2001] also found that the
improved Algorithm 9.3 outperformed Algorithm 9.2 as far as the patterns they
learned.
An Optimal Algorithm
As discussed in Section 9.1.1, a problem with a greedy search algorithm is
that it could halt at a candidate solution that locally maximizes the objective
function rather than globally maximizes it, and we can deal with this problem
using techniques such as iterated hill-climbing. However, these techniques only
improve our search; they do not guarantee we will find an optimal solution.
In 1997 Meek developed an algorithm called GES (Greedy Equivalent Search)
which, if the size of the data set is sufficiently large, will definitely find a perfect
9.1. APPROXIMATE MODEL SELECTION
Operation
Insert X − Y
Insert X → Y
Delete X − Y
Delete X → Y
Reverse X → Y
Insert X → Z ← Y
525
Change in Score
+X
scoreB (d, Y, NX,Y
∪ PAY ) − scoreB (d, Y, NX,Y ∪ PAY )
scoreB (d, Y, ΩX,Y ∪ PA+X
Y ) − scoreB (d, Y, ΩX,Y ∪ PAY )
+X
scoreB (d, Y, NX,Y ∪ PAY ) − scoreB (d, Y, NX,Y
∪ PAY )
−X
scoreB (d, Y, NY ∪ PAY ) − scoreB (d, Y, NY ∪ PAY )
+Y
scoreB (d, Y, PA−X
Y ) + scoreB (d, X, PAX ∪ ΩY,X )
−scoreB (d, Y, PAY ) − scoreB (d, X, PAX ∪ ΩY,X )
−Z+X
−Z
scoreB (d, Z, NX,Y
∪ PA+Y
Z ) + scoreB (d, Y, NX,Y ∪ PAY )
−Z+X
−scoreB (d, Z, NX,Y
∪ PAZ ) + scoreB (d, Y, NX,Y ∪ PAY )
Table 9.4: Local change in score for each operation in PDAGOPS.
Data set
MS Web
Nielsen
Each Movie
Media Metrix
House Votes
Mushroom
Time
Ratio
.95
.96
2.90
1.49
1.27
2.81
Table 9.5: A comparison of time usage of Algorithms 8.2 and improved Algorithm 8.3.
map if one exists. In 2002 Chickering proved this is the case. We describe the
algorithm next.
In what follows we denote the equivalence class represented by DAG pattern
gp by gp. GES is a two phase algorithm that searches over DAG patterns.
In the first phase, DAG pattern gp0 is in the neighborhood of DAG pattern
gp, denoted Nbhd+ (gp), if there is some DAG G ∈ gp for which a single edge
addition results in a DAG G0 ∈ gp0 . Starting with the DAG pattern containing
no edges, we repeatedly replace the current DAG pattern gp by the DAG pattern
in Nbhd+ (gp) which has the highest score of all DAG patterns in Nbhd+ (gp).
We do this until there is no DAG pattern in Nbhd+ (gp) which increases the
score.
The second phase is completely analogous to the first phase. In this phase,
DAG pattern gp0 is in the neighborhood of DAG pattern gp, denoted Nbhd− (gp),
if there is some DAG G ∈ gp for which a single edge deletion results in a
DAG G0 ∈ gp0 . Starting with the DAG pattern obtained in the first phase,
we repeatedly replace the current DAG pattern gp by the DAG pattern in
Nbhd− (gp) which has the highest score of all DAG patterns in Nbhd− (gp). We
do this until there is no DAG pattern in Nbhd− (gp) which increases the score.
Proof of Optimality We do not show an algorithm for GES since its structure is analogous to that of Algorithm 9.3. Rather we prove its optimality via
526 CHAPTER 9. APPROXIMATE BAYESIAN STRUCTURE LEARNING
the two theorems that follow. First we need a definition:
Definition 9.1 Let P be a joint probability distribution of the random variables
in a set V. If for every variables X ∈ V and every two subsets S, T ⊆ V , whenever IP ({X}, S|T) is not the case, there exists a Y ∈ S such that IP ({X}, {Y }|T)
is not the case, then we say the composition property holds for P .
Theorem 9.1 Given a class of augmented Bayesian networks for which size
equivalence holds, if we have a consistent scoring criterion for DAGs, and if
the generative distribution P satisfies the composition property, then for M (the
sample size) sufficiently large, the first phase of GES yields an independence
map of P .
Proof. The proof can be found in [Chickering and Meek, 2002].
The proof that phase one of the GES algorithm yields an independence
map assumes the composition property holds for the probability distribution.
This property holds for many important probability distributions. Namely,
Chickering and Meek [2002] show it holds for all probability distributions which
admit embedded faithful DAG representations. They show further that it holds
in the case of selection bias even if the observed probability distribution does
not admit an embedded faithful DAG representation. Recall Figure 2.33 which
appears again as Figure 9.4. In Exercise 2.35, we discussed that if the probability
distribution of X, Y , Z, W , and S is faithful to the DAG in Figure 9.4, then the
probability distribution of X, Y , Z, and W conditional on S = s does not admit
an embedded faithful DAG representation. This situation might happen when
selection bias is present (See Section 1.4.1.). Chickering and Meek [2002] show
the composition property holds for such conditional distributions. Specifically,
they show that if we have a probability distribution of a set of variables W which
admits a faithful DAG representation, V ⊆ W, O ⊆ V, and S = V − O, then the
composition property holds for the probability distribution of O conditional on
S = s.
Theorem 9.2 Given a class of augmented Bayesian networks for which size
equivalence holds, if we have a consistent scoring criterion for DAGs, and the
second phase of GES starts with an independence map of the generative distribution P , then for M (the sample size) sufficiently large, the second phase yields
an independence inclusion optimal map of P .
Proof. The proof is similar to that of Theorem 8.7 and is left as an exercise.
Note that for the second phase to yield an independence inclusion optimal
map, it can start with any independence map. Therefore, it could simply start
with the complete DAG pattern. The problem of course with the complete
DAG pattern is that in general it would require a prohibitively large number of
parameters. The purpose of the first phase is that hopefully it will identify a
simple independence map.
Theorem 9.3 Given a class of augmented Bayesian networks for which size
equivalence holds, if we have a consistent scoring criterion for DAGs, and if the
9.1. APPROXIMATE MODEL SELECTION
527
Y
Z
X
W
S
Figure 9.4: Selection bias is present.
generative distribution P admits a faithful DAG representation, then for M (the
sample size) sufficiently large, GES yields the DAG pattern which is a perfect
map of P .
Proof. The proof follows from the previous theorems, and the fact that if P admits a faithful DAG representation, then P satisfies the composition property.
Time Complexity Chickering [2002] developed a method for efficiently generating the members of Nbhd(gp)and for updating the score locally. The updating is similar to that shown in Section 9.1.2. In the first phase of the algorithm,
at most θ(n2 ) edges can be added and in the second phase at most θ(n2 ) edges
can be deleted, where n is the number of nodes. So in terms of states visited we
have a quadratic time algorithm. Recall from Section 9.1 Chickering [1996a] has
proven that for certain classes of prior distributions the problem of finding the
most probable DAG pattern is NP-complete. Therefore, it is unlikely we have a
polynomial-time algorithm. The nonpolynomial-time behavior of this algorithm
is in the possible number of members of Nbhd(gp) which must be investigated
in each step. For example, if in gp there is an edge between X and Y and Y has
k neighbors which are not adjacent to X, there may be at least θ(2k ) members
of Nbhd(gp). In experiments performed by Chickering [2002] the density of the
search space was never found to be a problem. If it did become a problem in
some real-world application, a heuristically-selected subset of Nbhd(gp) could
be investigated at each step; however, the large-sample optimality guarantee
would be lost in this case.
Comparison to Algorithms 9.2 and 9.3 Chickering [2002] compared the
GES algorithm to Algorithm 9.2 and the improved Algorithm 9.3. When 100
different gold standard Bayesian networks were used to generate data sets, GES
learned the equivalence classes to which the gold standards belong much more
528 CHAPTER 9. APPROXIMATE BAYESIAN STRUCTURE LEARNING
Y
Z
Y
Z
X
W
X
W
(a)
(b)
Figure 9.5: Two independence inclusion optimal maps of P |s, when P is the
faithful to the DAG in Figure 9.4 and all variables are binary except for X,
which has space size 4. Only the one in (a) is parameter optimal.
often than Algorithm 9.2 or 9.3. On the other hand, when the algorithms were
applied to the six real-life data sets shown in 9.5, there was no significant difference in the scores of the DAG patterns learned by Algorithm 9.3 and GES.
Indeed, in the case of Media Metrix and Mushroom, the algorithms learned the
exact same pattern, which contained no directed links. In all cases the patterns learned were reasonably sparse. Since in these the experiments the local
maximums can be identified with applying many operators that create uncoupled head-to-head meetings, the algorithms essentially traversed the same set
of states. Chickering [2002] suspects that in domains where there are more
complicated dependencies, the GES algorithm will identify different patterns,
and, given the results using the gold standards, better patterns. Finally, Chickering [2002] suggests using the GES algorithm as follows. First run a simple,
fast DAG-based greedy algorithm like Algorithm 9.2. If the resultant pattern
is sparse and contains few compelled edges, a more sophisticated algorithm will
probably not find a better solution. On the other hand, if the model is reasonably complicated, try the GES algorithm.
Are There Other Optimality Properties? Given that the GES algorithm
does terminate in a reasonable amount of time without resorting to some heuristic, there are two things that could prevent us from finding an optimal solution.
The first is that the sample size is not sufficiently large. There is nothing we
can do about this other than find a larger data set. The second is that the
generative distribution P does not admit a faithful DAG representation. An
interesting question is whether the GES algorithm has any other optimal properties in this case. We know that, for a sufficiently large data set, it finds an
independence inclusion optimal map. However, does it also find a parameter
optimal map? Chickering and Meek [2002] ran an experiment indicating it does
not. We describe the results of their experiment next.
Suppose P is faithful to the DAG in Figure 9.4, and all variables are binary
9.1. APPROXIMATE MODEL SELECTION
529
except for X, which has space size 4. Representative DAGs from the two DAG
patterns, which are inclusion optimal maps of P |s, appear in Figure 9.5. The
Bayesian network containing the DAG in Figure 9.5 (a) has 19 parameters and
is a parameter optimal map of P |s, while the Bayesian network containing the
DAG in Figure 9.5 (b) contains 21 parameters and is not. Owing to size equivalence, every DAG in the equivalence class containing the DAG in Figure 9.5 (a)
yields a Bayesian network containing 19 parameters. Similarly, those containing
DAGs from the equivalence class containing the DAG in Figure 9.5 (b) yield networks containing 21 parameters. Chickering and Meek [2002] generated many
Bayesian networks (each time with different a probability distribution) containing the DAG in Figure 9.4 and with all variables binary except for X, which has
space size 4. For each network, they instantiated S to its first value, generated
large data sets, and used the GES algorithm to learn a DAG pattern containing
the other four variables. They found that the GES algorithm learned each DAG
pattern about half the time, indicating a parameter optimal map is not always
learned when the data set is sufficiently large.
The experiment was conducting using a probability distribution which does
not admit an embedded faithful DAG representation. If we assume embedded
faithfulness, perhaps the GES algorithm does always identify a parameter optimal map (for sufficiently large data sets). However, this seems unlikely since
it is not hard to create a situation similar to the one used in Chickering and
Meek’s [2002] experiment, but in which P does admit an embedded faithful DAG
representation. Another possibility is that, even though GES does not always
identify a parameter optimal map, it might always identify an independence
inclusion optimal map containing the maximal number of conditional independencies. That is, there is no independence map containing more conditional
independencies. Future research might investigate these possibilities.
9.1.3
An Algorithm Assuming Missing Data or Hidden
Variables
The algorithms developed in the previous two subsections can be used when we
have missing data or hidden variables. In this case we would use as our scoring
function an approximate value obtained using one of the methods in Section
8.3.1 or 8.3.2. However, we have a problem in that with these scoring functions
we no longer have local scoring updating. So we must do a new expensive
computation for each candidate model we consider, which means we must do
many expensive computations before we can make a single change in the model.
Next we present the Structural EM Algorithm [Friedman, 1998]which only
needs to do an expensive computation each time we change the model. We
show the algorithm as a modification of Algorithm 9.2, which means it searches
for DAGs. However, it is straightforward to also modify Algorithm 9.3 and
search for DAG patterns. Furthermore, the method holds for both discrete and
continuous variables, but for simplicity we develop it using notation for discrete
variables.
The algorithm starts with some initial DAG G (say the empty DAG), and
530 CHAPTER 9. APPROXIMATE BAYESIAN STRUCTURE LEARNING
it computes (actually estimate using Algorithm 6.1) the MAP value f̃ (G) of f (G)
relative to the data. That is, it determines the value of f (G) that maximizes
ρ(f (G) |d, G). For a given Xi and parent set PA the algorithm then scores as
follows (See the discussion following Equality 9.3 for a discussion of the notation
in this formula.):
q(PA)
score0B (d, Xi , PA)
=
Y
j=1
where
0(PA)
sijk
P (PA)
(PA)
0(PA)
ri
Y
Γ( k aijk )
Γ(aijk + sijk )
, (9.5)
P (PA) P 0(PA)
(PA)
Γ( k aijk + k sijk ) k=1
Γ(aijk )
(G)
= E(s(PA)
)=
ijk |d, f̃
M
X
h=1
P (Xi(h) = k, paj |x(h) , f̃ (G) ).
That is, the calculation is the same as the one used in Algorithm 6.1 (EM MAP
determination) except the probability distribution used to compute the expected
value is the MAP value of the parameter set for G. The algorithm scores each
candidate DAG G00 as the product of the scores of its variables paired with
their parents in G00 . Using this scoring methodology, it proceeds exactly as in
Algorithm 9.2 except in each iteration it scores using the MAP value f̃ (G) where
G is the DAG chosen in the previous iteration.
Algorithm 9.4 Structural EM
Problem: Find a DAG G that approximates maximizing score0B (d, G) as given
by Equality 9.5.
Inputs: A Bayesian network structure learning schema BL containing n variables, data d with missing values or hidden variables.
Outputs: A set of edges E in a DAG that approximates maximizing score0B (d, G).
void Struct_EM (Bayes_net_struct_learn_schema BL,
data d,
set_of_edges& E)
{
E = ∅; G = (V, E);
do
// This is MAP Value.
(G)
(G)
= value that maximizes ρ(f |d, G); // Use Algorithm 6.1.
f̃
compute score0B (d, G);
if (any application of an operation in DAGOPS increases score0B (d, G))
modify E according to the one that increases score0B (d, G) the most;
while (some operation increases score0B (d, G));
}
9.2. APPROXIMATE MODEL AVERAGING
531
The preceding algorithm is intuitively appealing for two reasons: 1) if there
is no missing data, it reduce to Algorithm 8.2; and 2) it computes the score
using expected values relative to the MAP value of the parameter set for the
most recent best DAG. However, is there a theoretical justification for it? We
discuss such a justification next. Suppose we score each variable and parent set
pairing using


(PA)
qY
E [ln(scoreB (d, Xi , PA))] = E ln 
j=1
(PA)
Γ(Nj
(PA)
Γ(Nj
)
(PA)
+ Mj
(PA)
(PA)
ri
Y
Γ(aijk + sijk )
) k=1
(PA)
Γ(aijk )

 ,
where the expected value is again relative to the MAP value of the parameter set
for the DAG from the previous iteration. The score of a DAG is then the sum of
the scores of the variables paired with their parents in the DAG. Friedman [1998]
shows that if we use this scoring methodology in Algorithm 9.4, we are approximating a method that improves the actual score (probability of the data given
the model) in each step. A simple approximation of E [ln(scoreB (d, Xi , PAi ))]
is ln(score0 (Xi , PAi )). That is, we move the expectation inside the logarithm.
This approximation is what we used in Algorithm 9.4. Friedman [1998] shows
Equality 9.5 is a good approximation when the expected counts are far from 0.
Furthermore, Friedman [1998] develops more complex approximations for cases
where Equality 9.5 is not appropriate.
9.2
Approximate Model Averaging
As mentioned in Section 8.2, Heckerman et al [1999] illustrate that when the
number of variables is small and the amount of data is large, one structure can
be orders of magnitude more likely than any other. In such cases approximate
model selection yields good results. However, if the amount of data is not
large, it seems more appropriate to do inference by averaging over models as
illustrated in the Example 8.2. Another application of model averaging would
be to learn partial structure when the amount of data is small relative to the
number of variables. For example, Friedman et al [2000] discuss learning the
mechanism by which genes in a cell produce proteins, which then cause other
genes to express themselves. In this case, there are thousands of genes, but
typically we have only a few hundred data items. In such cases there are often
many structures which are equally likely. So choosing one particular structure
is somewhat arbitrary. However, in these cases we are not always interested
in learning the entire structure. That is, rather than needing the structure for
inference and decision making, we are only interested in learning relationships
among some of the variables. In particular, in the gene expression example, we
are interested in the dependence and causal relationships between the expression
levels of certain genes (See [Lander, 1999].) Model averaging would be useful
in this case. That is, by averaging over the highly probable structures, we may
532 CHAPTER 9. APPROXIMATE BAYESIAN STRUCTURE LEARNING
X1
X2
X3
X1
X2
X3
X1
X2
X3
X1
X3
X2
X1
X2
X3
X1
X3
X2
X2
X1
X3
X1
X2
X3
X1
X3
X2
X2
X1
X3
X1
X2
X3
Figure 9.6: The 11 DAG patterns when there are 3 nodes.
reliably learn there is a dependence between the expression levels of certain
genes.
Madigan and Raftery [1994] developed an algorithm called Occam’s Window, which does approximate model averaging. Madigan and York [1995] developed another approach which uses the Markov Chain Monte Carlo (MCMC)
method. That is the approach discussed here.
After presenting an example of using model averaging to learn partial structure, we develop MCMC algorithms for approximate model averaging.
9.2.1
A Model Averaging Example
Example 8.6 illustrated using model averaging to do inference. The following
example illustrates using it to learn partial structure (as in the gene expression
application discussed above).
Example 9.2 Suppose we have 3 random variables X1 , X2 , and X3 . Then the
possible DAG patterns are the ones in Figure 9.6. We may be interested in the
probability that a feature of the DAG pattern is present. For example, we may
be interested in the probability that there is an edge between X1 and X2 . Given
the 5 DAG patterns in which there is an edge, this probability is 1, and given
9.2. APPROXIMATE MODEL AVERAGING
533
the 6 DAG pattern in which there is no edge, this probability is 0. In general if,
we let F be a random variable whose value is present if a feature is present,
X
P (F = present|gp, d)P (gp|d)
P (F = present|d) =
gp
=
X
P (F = present|gp)P (gp|d),
gp
where
P (F = present|gp) =
½
1
0
if the feature is present in gp
if the feature is not present in gp.
You may wonder what event a feature represents. For example, what event
does an edge between X1 and X2 represent? This event is the event that X1
and X2 are not independent and are not conditionally independent given X3
in the actual relative frequency distribution of the variables. Another possible
feature is that there is a directed edge from X1 to X2 . This feature is the
event that, assuming the relative frequency distribution admits a faithful DAG
representation, there is a directed edge from X1 to X2 in the DAG pattern
faithful to that distribution. Similarly, the feature there is a directed path from
X1 to X2 represents the event there is a directed path from X1 to X2 in the
DAG pattern faithful to that distribution. Given we are only discussing the
relative frequency distribution, these events are ordinarily not of great interest.
However, if we are discussing causality, they tell us something about the causal
relationships among the variables. This matter is discussed in Section 10.2.4.
9.2.2
Approximate Model Averaging Using MCMC
As mentioned at the beginning of this section, when the number of possible
structures is large, we cannot average over all structures. In these situations
we heuristically search for high probability structures, and then we average
over them. Next we discuss doing this using the Markov Chain Monte Carlo
(MCMC) method.
Recall our two examples of model averaging (Examples 8.6 and 9.2). The first
involved computing a conditional probability overall all possible DAG patterns.
That is, we wish to compute
X
P (a|b, gp, d)P (gp|a, d).
P (a|b, d) =
gp
The second involved computing the probability a feature is present as follows:
X
P (F = present|d) =
P (F = present|gp)P (gp|d).
gp
In general, these problems involve some function of the DAG pattern and possibly the data, and a probability distribution of the patterns conditional on the
534 CHAPTER 9. APPROXIMATE BAYESIAN STRUCTURE LEARNING
data. So we can represent the general problem to be the determination of
X
f(gp, d)P (gp|d),
(9.6)
gp
where f is some function of gp and possibly d, and P is some probability distribution of the DAG patterns.
To approximate the value of Expression 9.6 using MCMC, our stationary
distribution r is P (gp|d). Ordinarily we can compute P (d|gp) but not P (gp|d).
However, if we assume the prior probability P (gp) is the same for all DAG
patterns,
P (d|gp)P (gp)
P (d)
= kP (d|gp)P (gp),
P (gp|d) =
where k does not depend on gp. If we use Equality 8.16 or 8.17 as our expression
for α, k cancels out of the expression, which means we can use P (d|gp) in the
expression for α. Note that we do not have to assign equal prior probabilities
to all DAG patterns. That is, we could use P (d|gp)P (gp) in the expression for
α also.
If we average over DAGs instead of DAG patterns, the problem is the determination of
X
f(G)P (G|d),
G
where f is some function of G and P is some probability distribution of the
DAGs. As is the case for DAG patterns, if we assume the prior probability P (G)
is the same for all DAGs, then P (G|d) = kP (d|G), and we can use P (d|G) in
the expression for α. However, we must realize what this assumption entails.
If we assign equal prior probabilities to all DAGs, DAG patterns containing
more DAGs will have higher prior probability. As noted in Section 9.1.2, when
performing model selection, assigning equal prior probabilities to DAGs is not
necessarily a serious problem as we will finally select a high-scoring DAG which
corresponds to a high-scoring DAG pattern. However, when performing model
averaging, a given DAG pattern will be included in the average according to
the number of DAGs in the pattern. For example, there are three DAGs corresponding to the DAG pattern X − Y − Z but only one corresponding to DAG
pattern X → Y ← Z. So by assuming all DAGs have the same prior probability,
we are assuming the prior probability that the actual relative frequency distribution has the set of conditional independencies {IP (X, Z|Y )} is three times the
prior probability that it has the set of conditional independencies {IP (X, Z)}.
Even more dramatic, there are n! DAGs corresponding to the complete DAG
pattern and only one corresponding to the DAG pattern with no edges. So we
are assuming the prior probability that are no conditional independencies is far
greater than the prior probability that the variables are mutually independent.
This assumption has consequences as follows: Suppose, for example, there are
9.2. APPROXIMATE MODEL AVERAGING
X1
X2
(a)
X3
X1
535
X2
X3
(b)
Figure 9.7: These DAGs are in each others neighborhoods, but their neighborhoods do not contain the same number of elements.
two variables X and Y , the correct DAG pattern is the one without an edge,
and we are interested in the feature IP (X, Y ). If we average over DAGs and
assign all DAGs equal prior probability, the posterior probability of the pattern
X − Y will be larger than it would be if we averaged over patterns and assigned
all patterns equal prior probability. Therefore, the feature is less probable if we
averaged over DAGs than if we averaged over patterns, which means averaging
over patterns has a better result. On the other hand, if the correct pattern is
X −Y and the feature of interest is qIP (X, Y ), we will confirm the feature more if
we average over DAGs. So averaging over DAGs has a better result. We see then
that we need look at the ensemble of all relative frequency distributions rather
than any one to discuss which method may be ‘correct’. If relative frequency
distributions are distributed uniformly in nature and we assign equal prior probabilities to all DAG patterns, then P (F = present|d), obtained by averaging
over DAG patterns, is the relative frequency with which we are investigating
a relative frequency distribution with this feature when we are observing these
data. So averaging over DAG patterns is ‘correct’. On the other hand, if relative frequency distributions are distributed in nature according to the number
of DAGs in DAG patterns and we assign equal prior probabilities to all DAGs,
then P (F = present|d), obtained by averaging over DAGs, is the relative frequency with which we are investigating a relative frequency distribution with
this feature when we are observing these data. So averaging over DAGs is ‘correct’. Although it seems reasonable to assume relative frequency distributions
are distributed uniformly in nature, some feel a relative frequency distribution,
represented by a DAG pattern containing a larger number of DAGs, may occur
more often because there are more causal relationships that can give rise to it.
After developing methods for averaging over DAGs, we develop a method
for averaging over DAG patterns.
Averaging over DAGs
A Straightforward Algorithm We show how to use MCMC to approximate
averaging over DAGs. Our set of states is the set of all possible DAGs containing
the variables in the application, and our stationary distribution is P (G|d), but
as noted previously, we can use P (d|G) in our expression for α. Recall from
Section 9.1.1 that Nbhd(G) is the set of all DAGs which differ from G by one
edge addition, one edge deletion, or one edge reversal. Clearly Gj ∈ Nbhd(Gi )
536 CHAPTER 9. APPROXIMATE BAYESIAN STRUCTURE LEARNING
if and only if Gi ∈ Nbhd(Gj ). However, since adding or reversing an edge can
create a cycle, if Gj ∈ Nbhd(Gi ) it is not necessarily true that Nbhd(Gi ) and
Nbhd(Gj ) contain the same number of elements. For example, if Gi and Gj are
the DAGs in Figures 9.7 (a) and (b) respectively, then Gj ∈ Nbhd(Gi ). However,
Nbhd(Gi ) contains 5 elements because adding the edge X3 → X1 would create a
cycle, whereas Nbhd(Gj ) contains 6 elements. We create our transition matrix
Q as follows: For each pair of states Gi and Gj we set

1

Gj ∈ Nbhd(Gi )

|Nbhd(Gi )|
,
qij =


/ Nbhd(Gi )
0
Gj ∈
where |Nbhd(Gi )| returns the number of elements in the set. Since Q is not
symmetric, we use Equality 8.16 rather than Equality 8.17 to compute αij .
Specifically, our steps are as follows:
1. If the DAG at the kth trial is Gi choose a DAG uniformly from Nbhd(Gi ).
Suppose that DAG is Gj .
2. Choose the DAG for the (k + 1)st trial to be Gj with probability

P (d|Gj ) × |Nbhd(Gi )|


1
≥1



P (d|Gi ) × |Nbhd(Gj )|
αij =
,


)
|Nbhd(G
)|
)
×
|Nbhd(G
)|
P
(d|G
P
(d|G

j
i
j
i

≤1

P (d|Gi ) |Nbhd(Gj )|
P (d|Gi ) × |Nbhd(Gj )|
and to be Gi with probability 1 − αij .
A Simplification It is burdensome to compute the sizes of the neighborhoods
of the DAGs in each step. Alternatively, we could include DAGs with cycles in
the neighborhoods. That is, Nbhd(Gi ) is the set of all graphs (including ones
with cycles) which differ from Gi by one edge addition, one edge deletion, or
one edge reversal. It is not hard to see that then the size of every neighborhood
is equal n(n − 1). We therefore define

1

Gj ∈ Nbhd(Gi )

n(n − 1)
.
qij =


/ Nbhd(Gi )
0
Gj ∈
If we are currently in state Gi and we obtain a graph Gj which is not a DAG,
we set P (d|Gj ) = 0 (effectively making rj zero). In this way αij is zero, the
graph is not chosen, and we stay at Gi in this step. Since Q is now symmetric,
we can use Equality 8.17 to compute αij . Notice that our theory was developed
assuming all values in the stationary distribution are positive, which is not
currently the case. However, Tierney [1995] shows convergence also follows if
we allow 0 values as discussed here.
9.2. APPROXIMATE MODEL AVERAGING
X1
X2
X3
X1
(a)
537
X2
X3
(b)
X1
X2
X3
X1
X2
X3
X1
X2
X3
X1
X2
X3
X1
X2
X3
X1
X2
X3
X1
X2
X3
X1
X2
X3
X1
X2
X3
(c)
(d)
Figure 9.8: The PDAGs in the neighorhood of the DAG patterns in (a) and (b)
are shown in (c) and (d) respectively.
Averaging over DAG Patterns
Next we develop an algorithm that approximates averaging over DAG patterns.
Recall from Section 9.1.2 that we developed the following complete set
PDAGOPS of operations for searching over DAG patterns:
1. If two nodes are not adjacent, add an edge between them that is undirected
or is directed in either direction.
2. If there is an undirected edge, delete it.
3. If there is a directed edge, delete it or reverse.
4. If there is an uncoupled undirected meeting Xi − Xj − Xk , add the uncoupled head-to-head meeting Xi → Xj ← Xk .
538 CHAPTER 9. APPROXIMATE BAYESIAN STRUCTURE LEARNING
After applying one of these operations, our strategy was to find a consistent
DAG extension of the new PDAG if one exists, and then determine the DAG
pattern corresponding to the DAG extension. Recall further that the operations
in PDAGOPS do not always yield a PDAG which results in a new DAG pattern
being chosen. Sometimes an operation yields a PDAG which does not admit a
consistent extension; and other times an edge insertion yields a PDAG which
admits a consistent extension, but the edge does not have the same direction in
the resultant DAG pattern as it had when inserted, which we do not allow. In
Section 9.1.2 we called Nbhd(gp) the set of all allowable DAG patterns which
can be obtained from DAG pattern gp. The determination of the sizes of these
neighborhoods is not straightforward. For the purpose of model averaging, we
can instead let Nbhd(gp) be the set of all PDAGs that result from applying one
of the operations in PDAGOPS to gp. However, Nbhd(gp) will then contain
some PDAGs that do not result in a new DAG pattern being chosen. We can
solve this problem using the simplified method shown in Section 9.2.2. That is,
if we obtain a PDAG gp0 in Nbhd(gp) that does not result in a DAG pattern
being chosen, we stay at gp in this step. Note that, due to our restriction
concerning edge direction, each DAG pattern obtainable from gp is chosen in
only one way. So if we choose uniformly from all PDAGs in Nbhd(gp), we will be
choosing each such DAG pattern with probability 1/ |Nbhd(gp)|. The situation
is not as simple as was in Section 9.2.2 because the transition Matrix Q is not
symmetric. For example, if gp is the DAG pattern in Figure 9.8 (a), Nbhd(gp)
is the set of PDAGs in Figure 9.8 (c). Yet if gp0 is the DAG pattern in Figure
9.8 (b), which clearly results from a member of Nbhd(gp), Nbhd(gp0 ) contains
the PDAGs in Figure 9.8 (d). We can solve this problem using the first method
shown in Section 9.2.2. That is, we count the number of PDAGs in Nbhd(gpi )
and use that value to determine the values of qij . In summary, we set
qij =





1
|Nbhd(gpi )|
0
gpj ∈ Nbhd(gpi )
,
/ Nbhd(gpi )
gpj ∈
and if we are currently in state gpi and obtain a PDAG gpj which does result
in a new DAG pattern being chosen, we stay at gpi in this step.
Next we discuss determining the size of the neighborhoods. We can do that
fairly efficiently as follows: The reverse edge operation is not necessary for our
set of operations to be complete. First we remove this operation. Then the
number of possible operations on each edge is one regardless of the direction of
the edge (Namely to delete it.). For a given DAG pattern gpi let
m =
r =
number of edges
number of uncoupled undirected meetings.
Then
|Nbhd(gpi )| = 3 [n(n − 1)/2 − (m)] + m + r
= 3n(n − 1)/2 − 2m + r.
EXERCISES
539
If we start with the DAG pattern with no edges, initially m = r = 0. Each time
our next pattern results from an edge addition we add 1 to m, and each time it
results from an edge deletion we subtract 1 from m. Each time our next pattern
results from an edge addition or deletion, we determine if any uncoupled undirected meetings are added or eliminated, and we adjust r accordingly. Finally,
each time our next pattern results from an uncoupled head-to-head meeting insertion, we determine how uncoupled undirected meetings have been eliminated
(We know there is at least one.) and subtract that number from r. Note that
this bookkeeping can be done while we construct the DAG pattern, providing
little additional overhead. Alternatively, we could not remove the reverse edge
operation and perform a little more bookkeeping.
Madigan et al [1996] developed a method that averages over DAG patterns by
introducing an auxiliary variable that takes values in the set of total orderings
of the nodes. Friedman and Koller [2000] developed a method that averages
over orderings of the nodes instead of over DAGs or DAG patterns, found this
method to be well correlated with averaging over DAG patterns in the case
of small networks in which they were able to compute the averages exactly.
Furthermore, they found their method outperformed averaging over DAGs in
experiments involving recovering features from large data sets.
EXERCISES
Section 9.1
Exercise 9.1 Does the PDAG in Figure 9.9 admit a consistent extension? If
so, find one.
Exercise 9.2 Implement Algorithms 9.1, 9.2 and 9.3, and the GES Algorithm
in the programming language of your choice. Next develop a Bayesian network
in some domain in which you are familiar, use the Monte Carlo technique to
generate cases from that network, and finally use your algorithms to approximate
learning the most probable DAG pattern. Compare the DAG patterns learned
to the DAG pattern representing the Markov equivalence class to which your
original DAG belongs.
Section 9.2
Exercise 9.3 Assuming there are three variables X1 , X2 , and X3 , and all DAG
patterns have the same posterior probability (1/11), compute the probability of
the following features being present (assuming faithfulness):
1. Ip (X1 , X2 ).
540 CHAPTER 9. APPROXIMATE BAYESIAN STRUCTURE LEARNING
X1
X2
X3
X4
X5
Figure 9.9: A PDAG.
2. qIp (X1 , X2 ).
3. Ip (X1 , X2 |X3 ) and qIp (X1 , X2 ).
4. qIp (X1 , X2 |X3 ) and Ip (X1 , X2 ).
Exercise 9.4 Implement the algorithms for averaging over DAGs and for averaging over DAG patterns and compare their performance on various data sets.
Chapter 10
Constraint-Based Learning
In Chapters 8 and 9, we assumed we had a set of variables with an unknown
relative frequency distribution, and we developed methods for learning the DAG
structure from data by computing the probability of the data given different
DAG patterns. Here we take a different approach. Given the set of conditional
independencies in a probability distribution, we try to find a DAG for which
the Markov condition entails all and only those conditional independencies.
This is called constraint-based learning. Chapter 11 discusses the relative
advantages of the two approaches and shows how in practice they can be used
together.
We assume that we are able to determine (or at least estimate) the conditional independencies INDP in a probability distribution P . Usually we do
this from data. Since (owing to Lemma 2.2) we need concern ourselves only
with conditional independencies among disjoint sets, henceforth we will assume
INDP contains all and only these conditional independencies. Recall Theorem
2.1 which says that the Markov condition entails all and only those conditional
independencies that are identified by d-separation. So our goal is to find a DAG
whose d-separations are the same as INDP . As discussed at in Section 8.1.1,
since we cannot distinguish DAGs in the same Markov equivalence class from
the conditional independencies in P , we can only learn the DAG pattern whose
d-separations are the same as INDP from those conditional independencies. We
will use the term ‘probability’ rather than ‘relative frequency’ in this chapter
because, although in practice the conditional independencies ordinarily come
from a relative frequency distribution, this is not a necessity.
It is only possible to find a DAG pattern whose d-separations are the same
as INDP when P admits a faithful DAG representation. Section 10.1 develops algorithms for finding such a DAG pattern assuming P admits a faithful
DAG representation. In Section 10.2 this assumption is relaxed. Namely, in
that section it is assumed only that the set of conditional independencies can
be embedded faithfully in a DAG. Section 10.3 concerns obtaining the set of
conditional independencies from data. Finally, in Section 10.4 we discuss the
relationship between constraint-based learning and human causal reasoning.
541
542
CHAPTER 10. CONSTRAINT-BASED LEARNING
X
Y
Figure 10.1: Given the d-separation in Example 10.1, this must be the DAG
pattern.
10.1
Algorithms Assuming Faithfulness
Since we are looking for a DAG which has the same d-separations as a set of
conditional independencies, our search has nothing to do with numeric probability, which means the theory can be developed without involving the notion of
probability. Therefore, we obtain our results by discussing only DAG properties.
To do this we need to first define faithfulness without involving probability.
By a set of d-separations IND amongst a set of nodes V we mean a set
containing statements of d-separation for the members of V. For example,
IND = {I({X}, {Y }), I({X}, {Y }|{Z})}
(10.1)
is a set of d-separations amongst V = {X, Y, Z}. In general, such a set does
not entail that there actually is a DAG with the d-separations. Indeed, there
is no DAG containing the d-separations in Equality 10.1. Note that we do not
subscript the set or the d-separations in the set with a DAG because we do
know if there is a DAG with the d-separations. IND simply represents a set
of statements as show in Equality 10.1. We say a set of d-separations IND is
faithful to DAG G if IND contains all and only the d-separations in G and
therefore in all DAGs Markov equivalent to G. When IND is faithful to some
DAG, we say IND admits a faithful DAG representation. Recall that a
DAG pattern gp has the same links as every DAG in the Markov equivalence
class represented by the pattern, and Definition 2.8 says two sets of nodes are
d-separated in gp by a third set of nodes if they are d-separated by that set in
any (and therefore every) DAG G in the Markov equivalence class represented
by gp. When a set of d-separations IND admits a faithful DAG representation,
we say IND is faithful to the DAG pattern representing the DAGs with which
IND is faithful. Clearly gp is unique.
We show an algorithm that determines this unique DAG pattern from a set
of d-separations when the set admits a faithful DAG representation. First we
give some examples introducing the methodology used in the algorithm; then
we present the algorithm. After that, we develop an algorithm for determining whether a given set of d-separations admits a faithful DAG representation.
Finally, we apply our results to probability.
10.1.1
Simple Examples
Next we present several examples illustrating constraint-based learning.
10.1. ALGORITHMS ASSUMING FAITHFULNESS
X
543
Y
Figure 10.2: Given the d-separation in Example 10.2, this must be the DAG
pattern.
Example 10.1 Suppose V = {X, Y } and
IND = {I({X}, {Y })}.
Owing to Lemma 2.7, if IND admits a faithful DAG representation, the DAG
pattern faithful to IND must not have a link (edge without regard for direction).
So we can conclude only the DAG pattern in Figure 10.1 could be correct. It is
left as an exercise to show IND is faithful to this DAG pattern.
Example 10.2 Suppose V = {X, Y } and
IND = ∅.
Owing to Lemma 2.7, if IND admits a faithful DAG representation, the DAG
pattern faithful to IND must have the link X − Y . So we can conclude only the
DAG pattern in Figure 10.2 could be correct. It is left as an exercise to show
IND is faithful to this DAG pattern.
Example 10.3 Suppose V = {X, Y, Z} and
IND = {I({X}, {Y }|{Z})},
Owing to Lemma 2.7, if IND admits a faithful DAG representation, the DAG
pattern faithful to IND must have the links X − Z and Y − Z and only these
X
Y
X
Y
Z
Z
(a)
(b)
Figure 10.3: Given the d-separation in Example 10.4, the DAG pattern must
be the one in (a). Given the d-separation in Example 10.3, the DAG pattern
must be the one in (b).
544
CHAPTER 10. CONSTRAINT-BASED LEARNING
X
Y
X
Y
Z
Z
W
W
(a)
(b)
Figure 10.4: Given the d-separations in Example 10.5, the links must be as
shown in (a) and the DAG pattern must be the one in (b).
links. This means our DAG pattern can only be one of the patterns in Figure
10.3 (a) or (b). Since we have I({X}, {Y }|{Z}), Lemma 2.8 says the meeting
X − Z − Y must not be head-to-head. So we can conclude only the DAG pattern
in Figure 10.3 (a) could be correct. It is left as an exercise to show IND is
faithful to this DAG pattern.
Example 10.4 Suppose V = {X, Y, Z} and
IND = {I({X}, {Y })}.
Owing to Lemma 2.7, if IND admits a faithful DAG representation, the DAG
pattern faithful to IND must have the links X −Z and Y −Z and only these links.
This means our DAG pattern can only be one of the patterns in Figure 10.3 (a)
or (b). Since we have I({X}, {Y }), Lemma 2.8 says the meeting X − Z − Y
must be head-to-head. So we can conclude only the DAG pattern in Figure 10.3
(b) could be correct. It is left as an exercise to show IND is faithful to this DAG
pattern.
Example 10.5 Suppose V = {X, Y, Z, W } and our set IND of d-separations
contains all and only the d-separations entailed by the d-separations in the following set:
{I({X}, {Y }})
I({X, Y }, {W }|{Z})}.
For example, IND also contains the d-separation I({X}, {W }|{Z}) because it is
entailed by I({X, Y }, {W }|{Z}). Owing to Lemma 2.7, our DAG pattern must
have the links in Figure 10.4 (a). Since we have I({X}, {Y }), Lemma 2.8 says
the meeting X − Z − Y must be head-to-head. Since we have I({X}, {W }|{Z}),
Lemma 2.8 says the meeting X −Z −W must not be head-to-head. Since we have
already concluded the link X − Z must be X → Z, we can therefore conclude the
10.1. ALGORITHMS ASSUMING FAITHFULNESS
U
U
X
545
Y
X
Y
Z
Z
W
W
(a)
(b)
Figure 10.5: Given the d-separations in Example 10.6, the links must be as
shown in (a) and the DAG pattern must be the one in (b).
link Z − W must be Z → W . Figure 10.4 (b) shows the resultant DAG pattern.
It is left as an exercise to show IND is faithful to this DAG pattern.
Example 10.6 Suppose V = {X, Y, Z, U, W }, and our set IND of d-separations
contains all and only the d-separations entailed by the d-separations in the following set:
{I({X}, {Y }|{U })
I({U }, {Z, W }|{X, Y })
I({X, Y, U }, {W }|{Z})}.
Owing to Lemma 2.7, our DAG pattern must have the links in Figure 10.5 (a).
Since we have I({X}, {Y }|{U }), Lemma 2.8 says the meeting X − Z − Y must
not be head-to-head. The other links must all be directed as in Example 10.5
for the reasons given in that example. Figure 10.5 (b) shows the resultant DAG
pattern. It is left as an exercise to show IND is faithful to this DAG pattern.
10.1.2
Algorithms for Determining DAG patterns
Using the methodology illustrated in the examples in the previous subsection, we
next develop algorithms that determine a DAG pattern from the d-separations
in the pattern. After giving a basic algorithm, we improve it.
A Basic Algorithm
First we give the algorithm and prove its correctness. Then we give some examples applying it. Finally, we analyze it.
546
CHAPTER 10. CONSTRAINT-BASED LEARNING
Algorithm 10.1 Find DAG Pattern
Problem: Given a set IND of d-separations, determine the DAG pattern faithful to IND if there is one.
Inputs: a set V of nodes and a set IND of d-separations among subsets of the
nodes.
Outputs: If IND admits a faithful DAG representation, the DAG pattern gp
containing the d-separations in this set.
void f ind_DAG_pattern (set-of-nodes V,
set-of-d-separations IND,
graph& gp)
{
for (each pair of nodes X, Y ∈ V) {
// Use Algorithm 2.2.
search for a subset SXY ⊆ V
such that I({X}, {Y }|SXY );
if (no such set can be found) {
create the link X − Y in gp;
// Step 1
}
for (each uncoupled meeting X − Z − Y )
if (Z ∈
/ SXY )
orient X − Z − Y as X → Z ← Y ;
// Step 2
while (more edges can be oriented) {
for (each uncoupled meeting X → Z − Y );
orient Z − Y as Z → Y ;
// Step 3
for (each link X − Y
such that there is a path from X to Y )
orient X − Y as X → Y ;
// Step 4
}
for (each uncoupled meeting X − Z − Y
such that X → W , Y → W , and Z − W )
orient Z − W as Z → W ;
// Step 5
}
In an implementation of the algorithm we do not actually input a set of
d-separations. Rather they ordinarily come from a probability distribution.
Next we prove the correctness of the algorithm.
Lemma 10.1 If the set of d-separations, which are the input to Algorithm 10.1,
admit a faithful DAG representation, the algorithm creates a link between X
and Y if and only if there is a link between X and Y in the DAG pattern gp
containing the d-separations in this set.
10.1. ALGORITHMS ASSUMING FAITHFULNESS
X
Z
547
Y
W
Figure 10.6: The link Z − W must be Z → W .
Proof. The algorithm produces a link if and only if X and Y are not d-separated
by any in set in gp, which, owing to Lemma 2.7, is the case if and only if X
and Y are adjacent in gp.
Lemma 10.2 (soundness) If the set of d-separations, which are the input to
Algorithm 10.1, admit a faithful DAG representation, then any directed edge
created by the algorithm is a directed edge in the DAG pattern containing the
d-separations in this set.
Proof. We consider the directed edges created in each step in turn:
Step 2: The fact that such edges must be directed follows from Lemma 2.8.
Step 3: If the uncoupled meeting X → Z − Y were X → Z ← Y , Z would not
be in any set that d-separates X and Y due to Lemma 2.8, which means
we would have created the orientation X → Z ← Y in Step 2. Therefore,
X → Z − Y must be X → Z → Y .
Step 4: If X − Y were X ← Y , we would have a directed cycle. Therefore, it
must be X → Y .
Step 5: The situation in Step 5 is depicted in Figure 10.6. If Z − W were
Z ← W , then X − Z − Y would have to be X → Z ← Y because otherwise
we would have a directed cycle. But if this were the case, we would have
created the orientation X → Z ← Y in Step 2 (See the analysis of Step
3.). So Z − W must be Z → W .
Lemma 10.3 (completeness) If the set of d-separations, which are the input
to Algorithm 10.1, admit a faithful DAG representation, all the directed edges,
in the DAG pattern containing the d-separations in this set, are directed by the
algorithm.
Proof. The proof can be found in [Meek, 1995b].
Theorem 10.1 If the set of d-separations, which are the input to Algorithm
10.1, admit a faithful DAG representation, the algorithm creates the DAG pattern containing the d-separations in this set.
Proof. The proof follows from the preceding three lemmas.
548
CHAPTER 10. CONSTRAINT-BASED LEARNING
X
Y
Z
X
X
Y
U
Z
U
Y
Z
W
W
W
T
T
T
(a)
(b)
(c)
U
Figure 10.7: Given the d-separations in the DAG in (a), Algorithm 10.1 creates
the links in (b) and the directed edges in (c).
Some of the examples that follow concern cases in which IND is fairly large.
So rather than explicitly listing the d-separations in IND we show d-separations
using a DAG.
Example 10.7 Suppose V = {X, Y, Z, U, T, W } and IND consists of the dsepara-tions in the DAG in Figure 10.7 (a). Given this input, Algorithm 10.1
does the following:
Step 1 produces the links in Figure 10.7 (b).
Step 2 orients X − Z − Y as X → Z ← Y .
Step 3 orients Z–W as Z → W .
Step 4 orients X − W as X → W .
Step 3 orients W − T as W → T .
The resultant DAG pattern appears in Figure 10.7 (c).
Example 10.8 Suppose V = {X, Y, Z, T, W } and IND consists of the d-separations in the DAG in Figure 10.8 (a). Given this input, Algorithm 10.1 does the
following:
Step 1 produces the links in Figure 10.8 (b).
Step 2 orients X − W − Y as X → W ← Y .
10.1. ALGORITHMS ASSUMING FAITHFULNESS
X
Z
Y
X
Z
Y
549
X
Z
W
W
W
T
T
T
(a)
(b)
(c)
Y
Figure 10.8: Given the d-separations in the DAG in (a), Algorithm 10.1 produces
the links in (b) and the directed edges in (c).
Step 3 orients W − T as W → T .
Step 5 orients Z − W as Z → W .
The resultant DAG pattern appears in Figure 10.8 (c).
An implementation of Algorithm 10.1 could prove to be very inefficient because to determine whether there is a set d-separating X and Y in Step 1 we
might search all 2n−2 subsets of V not including X and Y . The algorithm given
next improves on this performance on the average.
An Improvement to the Algorithm
In the worse case, any algorithm would have to search all 2n−2 subsets of V −
{X, Y } to determine whether X and Y are d-separated. The reason is that
given a subset U ⊂ V − {X, Y }, X and Y could not be d-separated by U but
d-separated by some superset or subset of U. Therefore, in general we only
eliminate one set each time we test for d-separation. However, in the case of
sparse DAGs, we can improve things significantly by considering small subsets
first. Recall Corollary 2.2, which says if X and Y are d-separated in DAG G,
then either they are d-separated by the parents of X in G or by the parents of
Y in G. This means we need only consider subsets of ADJX in gp and subsets
of ADJY in gp when determining whether X and Y are d-separated in gp. By
ADJX we mean the subset of V consisting of all nodes adjacent to X. Next we
show an algorithm which does this. The algorithm looks first at subsets of size
0, then at subsets of size 1, then at subset of size 2, and so on.
First we give the algorithm, then we show an example, and finally we discuss
its efficiency. We call it the PC algorithm because that is the name given by its
developers in [Spirtes et al, 1993, 2000].
.
550
CHAPTER 10. CONSTRAINT-BASED LEARNING
Algorithm 10.2 PC Find DAG Pattern
Problem: Given a set IND of d-separations, determine the DAG pattern faithful to IND if there is one.
Inputs: a set V of nodes and a set IND of d-separations among subsets of the
nodes.
Outputs: If IND admits a faithful DAG representation, the DAG pattern gp
containing the d-separations in this set.
void P C (set-of-nodes V,
set-of-d-separations IND,
graph& gp)
{
form the complete undirected graph gp over V;
i = 0;
repeat
for (each X ∈ V)
for (each Y ∈ ADJX ) {
determine if there is a set S ⊆ ADJX − {Y }
such that |S| = i and I({X}, {Y }|S) ∈ IND;
if such a set S is found {
SXY = S;
remove the edge X − Y from gp;
}
}
i = i + 1;
until (|ADJX | < i for all X ∈ V);
for (each uncoupled meeting X − Z − Y )
if (Z ∈
/ SXY )
orient X − Z − Y as X → Z ← Y ;
while (more edges can be oriented) {
for (each uncoupled meeting X → Z − Y );
orient Z − Y as Z → Y ;
for (each X − Y
such that there is a path from X to Y )
orient X − Y as X → Y ;
for (each uncoupled meeting X − Z − Y
such that X → W , Y → W , and Z − W )
orient Z − W as Z → W ;
}
}
// | | returns
// size of set.
// Step 1
// Step 2
// Step 3
// Step 4
// Step 5
10.1. ALGORITHMS ASSUMING FAITHFULNESS
551
X
Y
X
Y
Z
W
Z
W
T
T
(a) a DAG
(b) the complete graph
X
Y
X
Y
Z
W
Z
W
T
T
(c) i = 0
(d) i = 1
X
Y
X
Y
Z
W
Z
W
T
T
(e) i = 2
(f) the final dag pattern
Figure 10.9: Given the d-separations in the DAG in (a), the steps in Algorithm
10.2 produce the graphs in (b)-(f).
552
CHAPTER 10. CONSTRAINT-BASED LEARNING
Before analyzing the preceding algorithm, we show an example of applying
it.
Example 10.9 Suppose V = {X, Y, Z, T, W } and IND consists of the d-separations in the DAG in Figure 10.9 (a). Given this input, Algorithm 10.2 does the
following:
Step 1 produces the complete undirected graph in Figure 10.9 (b)
Step 2 with i = 0 produces the undirected graph in Figure 10.9 (c);
Step 2 with i = 1 produces the undirected graph in Figure 10.9 (d);
Step 2 with i = 2 produces the undirected graph in Figure 10.9 (e);
Steps 3-6 produce the DAG pattern in Figure 10.9 (f).
Next we analyze the algorithm:
Analysis of Algorithm 10.2 (PC Find DAG Pattern)
We will determine a bound W (n) for the number of d-separation
tests required in the repeat-until loop. Let
n = number of nodes in V
k = maximum size of ADJX over all X ∈ V.
There are n choices for the value of X in the first for loop. Once X
is chosen, there are n − 1 choices
¢ Y . For given values of X, Y ,
¡ for
subsets for d-separating X and
and i, we must check at most n−2
i
Y . The value of i goes from 0 to at most k. So we have
W (n) ≤ n(n − 1)
¶
k µ
X
n−2
i=0
i
≤
n2 (n − 2)k
.
(k − 1)!
It is left as an exercise to derive the second inequality using induction
on k.
So we see, if the DAG pattern is sparse (i.e. no node is adjacent to very
many other nodes), the algorithm is reasonably efficient. Of course, in general
it is not.
10.1.3
Determining if a Set Admits a Faithful DAG Representation
Algorithms 10.1 and 10.2 do not check if the set of d-separations, which is
their input, admits a faithful DAG representation. They are just guaranteed to
produce the correct DAG pattern if it does. In Examples 10.7, 10.8, and 10.9
10.1. ALGORITHMS ASSUMING FAITHFULNESS
N
H
F
553
T
C
(a)
N
T
F
C
(b)
Figure 10.10: Given V = {N, F, C, T } and IND consists of the d-separations
in the DAG in (a) restricted to these variables, Algorithms 10.1 and 10.2 will
produce the graph in (b).
the d-separations come from a DAG; so we know they are the d-separations in
a DAG pattern. However, in general this is not the case. For example, in the
next subsection we will obtain the d-separations from a probability distribution.
So in general we need to check whether the output of Algorithms 10.1 and 10.2
is a DAG pattern. Next we discuss how to determine whether this is the case.
First we give examples in which it is not.
Example 10.10 Suppose V = {N, F, C, T } and IND consists of the d-separations in the DAG in Figure 10.10 (a) restricted to these variables. Given this
input, it is left as an exercise to show Algorithms 10.1 and 10.2 produce the
graph in Figure 10.10 (b). This graph is not a DAG pattern because it contains
a directed cycle. Theorem 10.1 says if the set of d-separations, which is the input
to Algorithm 10.1, admits a faithful DAG representation, the algorithm creates
that DAG pattern faithful to that set. Since the algorithm produced something
that is not a DAG pattern, we can conclude IND does not admit a faithful DAG
representation.
Note that Figure 10.10 (a) is the same as Figure 8.7 in Section 8.5.1. In that
section we said that the variable H is a hidden variable because it is not one of
the four variables on which we have data, and we shade nodes representing such
variables. Here we are using H only to facilitate showing a set of d-separations
amongst N , F , C and T , and there is no need to call it a hidden variable.
However, in Section 10.2 we will develop methods for learning a DAG pattern,
in which a set of d-separations amongst V is embedded faithfully, when the set
does not admit a faithful DAG representation. Using these methods, from the
554
CHAPTER 10. CONSTRAINT-BASED LEARNING
X
Y
X
Y
Z
Z
(a)
(b)
Figure 10.11: Given the d-separations in Example 10.11, Algorithm 10.2 will
produce the DAG pattern in (a) and Algorithm 10.1 could produce either DAG
pattern.
set IND in Example 10.10 we can learn a DAG pattern containing the DAG in
Figure 10.10 (a). This pattern will contain a variable other than those in V, and
it is hidden in the same way H is hidden in the situation discussed in Section
8.5.1. That is, it is not one of the four variables for which we have ‘data’. When
we apply the theory to probability, not having d-separations involving a variable
is the same as not having data on the variable. We shade the node now and call
it H for consistency.
Example 10.11 Suppose V = {X, Y, Z} and
IND = {I({X}, {Y }), I({X}, {Y }|{Z})}.
Given this input, Algorithm 10.2 produces the DAG pattern in Figure 10.11
(a), while Algorithm 10.1 produces either the one in Figure 10.11 (a) or the
one Figure 10.11 (b) depending on whether I({X}, {Y }) or I({X}, {Y }|{Z}) is
found first. It is left as an exercise to show neither DAG pattern contains both
d-separations. Owing to Theorem 10.1, we can conclude IND does not admit a
faithful DAG representation.
The previous example shows that, even if Algorithms 10.1 and 10.2 produce
a DAG pattern, we cannot assume this pattern contains the d-separations which
are the input to the algorithms. A straightforward way to determine whether
this is the case is to check whether the output DAG pattern has all and only
the input d-separations (as was done in Example 10.11). Next we develop an
algorithm for doing this.
Recall that a DAG pattern gp has the same d-separations as any DAG G
in the equivalence class represented by that DAG pattern. We can check if a
DAG G contains all and only a given set IND of d-separations as follows: It is
straightforward to check if every d-separation in G is in IND (We discuss how
to do this after presenting Algorithm 10.4.), and we can use Algorithm 2.2 in
Section 2.1.3 to determine whether every d-separation in IND is a d-separation in
G. So, if we first construct a DAG in the Markov equivalence class represented
10.1. ALGORITHMS ASSUMING FAITHFULNESS
555
by a given DAG pattern from that pattern, we can test whether the pattern
has a given set of d-separations. An algorithm for constructing such a DAG
follows:.
Algorithm 10.3
Construct DAG
Problem: Given a DAG pattern, construct any DAG in the equivalence class
represented by that pattern.
Inputs: a graph gp.
Outputs: if gp is a DAG pattern, a DAG G in the equivalence class represented
by gp.
void construct_DAG (graph gp,
graph& G)
{
while there are unoriented edges {
choose an unoriented edge X − Y ;
orient X − Y as X → Y ;
while (more edges can be oriented) {
if (X → Z − Y is an uncoupled meeting);
orient Z − Y as Z → Y ;
if (X − Y && there is a path from X to Y )
orient X − Y as X → Y ;
for (each uncoupled meeting X − Z − Y
such that X → W , Y → W , and Z − W )
orient Z − W as Z → W ;
for (each X → Y → Z -W − X such that
Y and W are linked, and Z and X are not linked)
orient W − Z as W → Z;
}
// Step 1
// Step 2
// Step 3
// Step 4
// Step 5
}
Theorem 10.2 If its input is a DAG pattern gp, Algorithm 10.3 produces a
DAG in the Markov equivalence class represented by gp.
Proof. The proof can be found in [Meek, 1995b].
Notice that the rules for orienting edges in Algorithm 10.3 are the same as
those in Algorithms 10.1 and 10.2 except that we have the additional rule in
Step 5. Let’s discuss that rule. The situation it concerns is shown in Figure
10.12. The dotted line means the link between Y and W could be Y − W ,
Y → W , or W → Y . Step 5 says that in this situation W − Z must be W → Z.
The reason is as follows. If we have W → Y , then Z → W would create a
556
CHAPTER 10. CONSTRAINT-BASED LEARNING
Z
Y
W
X
Figure 10.12: Given this situation, W − Z must be W → Z. The dotted line
means the link between Y and W could be Y − W , Y → W , or W → Y.
directed cycle. So we must have W → Z. If we have Y → W , then W → X
would create a directed cycle. So we must have X → W , and therefore, since
Z–W –X is an uncoupled meeting, we must have W → Z. You may wonder
why we do not need this rule in Algorithms 10.1 and 10.2. The reason is that,
when we start only with edges oriented from uncoupled meetings, the first three
steps always end up orienting W − Z as W → Z in this situation. It is left as
an exercise to show this.
It is not hard to modify Algorithm 10.3 so that it determines whether a
consistent extension of any PDAG exists and, if so, produces one. It is left as
an exercise to do so. Dor and Tarsi [1992] develop a quite different algorithm
which also accomplishes this, and which therefore could also be used for our
purposes here.
We have the following theorem, which enables us to create a DAG faithful
to IND when IND admits a faithful DAG representation.
Theorem 10.3 Suppose we have a set of d-separations which admits a faithful
DAG representation, and gp is the DAG pattern that has the d-separations in
this set. Then if this set is the input to Algorithm 10.1 or 10.2, and the output
of these algorithms is the input to Algorithm 10.3, this algorithm produces a
DAG which is a member of the Markov equivalence class represented by gp.
Proof. The proof follows immediately from Theorems 10.1 and 10.2.
We can now test whether a given set IND of d-separations admits a faithful
DAG representation as follows: We first use IND as the input to Algorithm
10.2. Then we use the output of Algorithm 10.2 as the input to Algorithm 10.3.
Finally, we test whether the output of Algorithm 10.3 is a DAG containing
all and only the d-separations in IND. If it is, we’ve shown there is a DAG
pattern containing all and only the d-separations in the set, while if it is not,
the preceding theorem implies there is no such DAG pattern. The algorithm
follows:
10.1. ALGORITHMS ASSUMING FAITHFULNESS
Algorithm 10.4
557
Determine Faithful
Problem: Determine whether a set of d-separations admits a faithful DAG
representation.
Inputs: A set of nodes V and a set of d-separations IND.
Outputs: A boolean variable switch and a graph G. If IND admits a faithful
DAG representation, swtch’s value is true and G is a DAG faithful to IND.
If IND does not, switch’s value is false.
void determine_f aithf ul (set-of-nodes V,
set-of-d-separations IND,
graph& G,
bool& switch)
{
graph gp;
switch = false;
P C(V, IND, gp);
construct_DAG(gp, G);
if (G does not contain a directed cycle)
if (set of d-separations in G ⊆ IND)
if (IND ⊆ set of d-separations in G)
switch = true;
// Call Algorithm 10.2.
// Call Algorithm 10.3.
// Use Algorithm 2.2
// to check this.
}
As discussed in Section 10.1.4, we are interested in the case where IND is
the set of conditional independencies INDP in a probability distribution P . In
this case, to determine whether every d-separation in G is in IND, we need only
check if the Markov condition is satisfied. That is, for each node X, if we denote
the sets of parents and nondescendents of X in G by PAX and NDX respectively,
we need to check whether
IP ({X}, NDX − PAX |PAX ).
The reason is that Lemma 2.1 then implies every d-separation in G is a conditional independency in P . It turns out we can even improve on this. Order
the nodes so that for each X all the ancestors of X in G are numbered before
X. Let RX be the set of nodes that precede X in this ordering. It is left as an
exercise to show that for each X we need only check whether
IP ({X}, RX − PAX |PAX ).
558
CHAPTER 10. CONSTRAINT-BASED LEARNING
T
T
X
Y
X
Z
Y
W
Z
W
(b) after Algorithm 10.2
(a) A DAG
T
T
X
X
Y
Z
W
(c) Step 1 chooses W
vY.
Y
Z
W
vX.
(d) Step 5 orients Z
T
T
X
X
Y
Z
W
(e) Step 1 chooses TvZ.
Y
Z
W
(f) Step 2 orients remaining edges.
Figure 10.13: Given the d-separations in the DAG in (a), the steps in Algorithm
10.4 are shown in (b)-(f)). The steps in (c)-(f) refer to those in Algorithm 10.3.
10.1. ALGORITHMS ASSUMING FAITHFULNESS
559
X
W
X
W
Y
Z
Y
Z
(a)
(b)
Figure 10.14: Given the d-separations in Example 10.13, Algorithm 10.4 will
first call Algorithm 10.2 to produce the graph in (a). After that, it will call
Algorithm 10.3 to produce the directed graph in (b) (or one with the arrows
reversed).
Example 10.12 Suppose V = {X, Y, Z, W, T } and IND consists of the d-separations in the DAG in Figure 10.13 (a). Given this as the input to Algorithm 10.4,
Algorithm 10.2 will produce the DAG pattern in Figure 10.13 (b). After that,
Algorithm 10.3 will do the steps shown in Figure 10.13 (c), (d), (e), and (f).
Of course, if other edges are chosen in Step 1 we could get a different DAG
(e.g. the one in (a)). The fact that the output of Algorithm 10.3 is a DAG does
not necessarily mean that IND admits a faithful DAG representation. Algorithm
10.4 still must check this. It is left as an exercise for you to do this using the
DAG in Figure 10.13 (f). Of course, we already knew it because we started with
the DAG in Figure 10.13 (a).
Example 10.13 Suppose V = {X, Y, Z, W } and
IND = {I({X}, {Z}|{Y, W }), I({Y }, {W }|{X, Z})}.
Given this as the input to Algorithm 10.4, it is left as an exercise to show that
Algorithm 10.2 will produce the graph in Figure 10.14 (a); and then Algorithm
10.3 will produce either the directed graph in Figure 10.14 (b) or one containing a
directed cycle with the direction of the arrows reversed. Since this graph contains
a directed cycle, it is not a DAG. So Theorem 10.3 implies there is no DAG
pattern containing d-separations in this set.
The previous example shows a case in which the output of Algorithm 10.4 is
not a DAG, which immediately enables us to conclude there is no DAG pattern
containing the d-separations. Example 10.10 also shows a case in which the
output of Algorithm 10.4 would not be a DAG. Example 10.11 shows a case
in which the output of Algorithm 10.4 would be a DAG, and the set of dseparations in the DAG is a subset of IND. However, IND is not a subset of
the d-separations in the DAG, which enables us to conclude there is no DAG
560
CHAPTER 10. CONSTRAINT-BASED LEARNING
H1
X
Y
H2
Z
W
T
W
T
(a)
X
Y
Z
(b)
Figure 10.15: Given V = {X, Y, Z, W, T } and IND consists of the d-separations
in the DAG in (a) restricted to these variables, Algorithm 10.4 will produce the
DAG in (b) or one with W → T instead of W ← T .
pattern containing the d-separations. The next example shows a situation in
which the output of Algorithm 10.4 is a DAG, but the set of d-separations in
the DAG is not a subset of IND.
Example 10.14 Suppose V = {X, Y, Z, W, T } and IND consists of the d-separations in the DAG in Figure 10.15 (a) restricted to these variables. Given this
input, it is left as an exercise to show Algorithm 10.4 will produce the DAG
in Figure 10.15 (b) or one with W → T instead of W ← T . It is further
left as an exercise to show that we have I({X}, {T }|{Y, Z, W }) in the DAG in
Figure 10.15 (b) but not in the one in Figure 10.15 (a). Therefore, the set of dseparations in the DAG in Figure 10.15 (b) is not a subset of IND, which means,
due to Theorem 10.3, IND does not admit a faithful DAG representation.
As in Figure 10.10 (a), nodes are shaded in Figure 10.15 (a) for consistency
with the figures in Section 10.2.
10.1.4
Application to Probability
Suppose we have a joint probability distribution P of the random variables in a
set V. Recall Theorem 2.6 says that if P admits a faithful DAG representation,
then the d-separations in the DAG pattern gp faithful to P identify all and
only conditional independencies in P . So assuming P admits a faithful DAG
representation, we can find the DAG pattern gp faithful to P by using the
10.2. ASSUMING ONLY EMBEDDED FAITHFULNESS
561
conditional independencies INDP in P as the input to Algorithms 10.1 and 10.2.
That is we do the following:
IND = INDP ;
P C(V, IND, gp);
Again in an implementation of the algorithm we do not actually input a set
of d-separations. This code means the d-separations are obtained from the
conditional independencies in the probability distribution.
Furthermore, to determine whether a given probability distribution P admits
a faithful DAG representation, we can call Algorithm 10.4 as follows:
IND = INDP ;
determine_pattern(V, IND, G, switch);
If P admits a faithful DAG representation, switch will be true and G will be
DAG faithful to P . If P does not, switch will be false.
Given the above, we can use Examples 10.10, 10.11, 10.13, and 10.14 to illustrate applications to probability. That is, suppose the d-separations in these
examples are the conditional independencies in a probability distribution P .
Then Examples 10.10 and 10.13 show cases in which we can conclude P does
not admit a faithful DAG representation because the output of Algorithm 10.4
is not a DAG. Recall a probability distribution, that has the conditional independencies which are the d-separations in Example 10.13, is shown in Exercise
2.19. Example 10.11 shows a case in which the output of Algorithm 10.4 is a
DAG, but still P does not admit a faithful DAG representation. Recall Figure
2.6 shows a probability distribution which has these conditional independencies,
and Example 2.13 shows that such a probability distribution does not even admit an embedded faithful DAG representation. Example 10.14 shows a case in
which P does admit an embedded faithful DAG representation, the output of
Algorithm 10.4 is a DAG, but yet P is not faithful to this DAG. Indeed, P does
not even satisfy the Markov condition with this DAG. This example shows that,
even if we assume P admits an embedded faithful DAG representation, the fact
that Algorithm 10.4 produces a DAG does not necessarily mean P admits a
faithful DAG representation.
10.2
Assuming Only Embedded Faithfulness
In Example 10.11, we noticed that Algorithm 10.1 does not even produce a
unique output. As noted above, a probability distribution, that has the conditional independencies which are the d-separations in that example, not only
does not admit a faithful DAG representation, it does not even admit an embedded faithful DAG representation. However, such distributions seem rare (See
the discussions of causation in Section 2.6 and of the results in [Meek, 1995a] in
Section 2.3.1.). Example 10.14 concerned a P which does not admit a faithful
DAG representation but does admit an embedded faithful DAG representation.
562
CHAPTER 10. CONSTRAINT-BASED LEARNING
In that example, we saw that, in this case, P does not necessarily even satisfy
the Markov condition with the DAG produced by Algorithm 10.4. Therefore,
we can conclude that, when we assume only embedded faithfulness, Algorithm
10.4 (and therefore Algorithms 10.1 and 10.2) cannot tell us much about P .
In this section, given the assumption of embedded faithfulness, we develop
algorithms for learning about the DAGs in which P is embedded faithfully.
Again we obtain our results by discussing DAGs. To do this we need to first
define embedded faithfulness without involving probability.
Let IND be a set of d-separations amongst the set V of nodes. We say IND is
embedded faithfully in DAG G if G = (W, E), V ⊆ W, and all and only the
d-separations in IND are d-separations in G restricted to elements of V. We say
IND admits an embedded faithful DAG representation if there is some
DAG in which IND is embedded faithfully. If IND admits an embedded faithful
DAG representation, it is embedded faithfully in an infinite number of DAGs.
We will develop algorithms that determine common features of every DAG
in which IND is embedded faithfully. After developing a basic algorithm, we
improve it. Then we apply our results to probability, and finally we show an
application to learning causal influences. Before we can do any of this, we need
to develop theory concerning inducing chains.
10.2.1
Inducing Chains
First we introduce some notation. Let G = (V, E) be a directed graph. Given
a chain ρ = [X1 , X2 , ..., Xk ] in G, we denote the subchain [Xi , Xi+1 , ..., Xj ] of ρ
between Xi and Xj by ρ(Xi , Xj ) and by ρ(Xj , Xi ). Given chains ρ = [X1 , . . . Xj ]
and γ = [Y1 , . . . Yk ] in G such that Xj = Y1 , the chain [X1 , . . . Xj , Y2 , . . . Yk ] is
called the concatenation of ρ and γ. We denote it as ρ ⊕ γ or γ ⊕ ρ. Let ρ be
a chain between X and Y in G. We say ρ is out of X if the edge touching X
on ρ has its tail at X; we say ρ is into X if the edge touching X on ρ has its
head at X. If Z is an interior node on ρ, and we have X → Z ← Y on ρ, then
Z is called a collider on ρ; and otherwise Z is called a non-collider on ρ.
A hidden node DAG is a DAG G = (V ∪ H, E), where V and H are disjoint
sets. The nodes in V are called observable nodes, while the nodes in H are
called hidden nodes. We denote members of H by H or Hi , while we denote
members of V by all other letters of the alphabet (e.g. X, Y , Z, etc). Note that
this definition corresponds to the definition of a hidden variable DAG model
given in Section 8.5. We say ‘hidden node’ instead of ‘hidden variable’ because
we are currently discussing only DAG properties. We shade hidden nodes.
Example 10.15 A hidden node DAG appears in Figure 10.16. In that DAG,
the chain ρ = [X, H1 , H2 , H3 , Y, Z] is out of Z and into X. The nodes Y and
H2 are colliders on ρ, while all other nodes on ρ are non-colliders.
We now have the following definition:
Definition 10.1 Suppose G = (V ∪ H, E) is a hidden node DAG. For X, Z ∈ V,
10.2. ASSUMING ONLY EMBEDDED FAITHFULNESS
563
T
X
H1
H2
H3
Y
Z
Figure 10.16: The chain [X, H1 , H2 , H3 , Y, Z] is an inducing chain over V in G
betwen X and Z, where V = {X, Y, Z, T }.
X=
6 Z, an inducing chain1 ρ over V in G between X and Z is a chain between
X and Z satisfying the following:
1. If Y ∈ V and Y is an interior node on ρ, then Y is a collider on ρ.
2. If W ∈ V ∪ H is a collider on ρ, then W is an ancestor in G of either X
or Z. That is, there is a path from W to either X or Z.
Example 10.16 Let H = {H1 , H2 , H3 } and V = {X, Y, Z, T }. Consider the
hidden node DAG G in Figure 10.16. The chain ρ = [X, H1 , H2 , H3 , Y, Z] is
an inducing chain over V in G between X and Z. We see that as follows: The
interior node Y is a collider on ρ, the collider Y is an ancestor in G of X, and
the collider H2 is an ancestor in G of Z.
Notice that it is not possible to d-separate X and Z using only a subset of
V. That is, if S ⊆ V is to d-separate X and Z, we need both T and Y ∈ S to
block the chains [X, H1 , H2 , T, Z] and [X, Y, Z] respectively. However including
T and Y in S unblocks the chain [X, H1 , H2 , H3 , Y, Z].
The definition of an inducing chain allows cycles. However, the following
lemma shows, when there is an inducing chain, we can assume there is one
without cycles.
Lemma 10.4 Suppose G = (V ∪ H, E) is a hidden node DAG, and X, Z ∈ V.
Suppose there is an inducing chain over V in G between X and Z which is
into/out of X and into/out of Z. Then there is a simple (one without cycles)
inducing chain between X and Z, which is into/out of X and into/out of Z.
Proof. Assume there is at least one inducing chain between X and Z, which is
into/out of X and into/out of Y . Let ρ = [V1 , . . . Vk ] be the one with the least
number m of cyclic subpaths. Note that X = V1 and Z = Vk . If m = 0, we are
done. Otherwise, there are some i and j, i 6= j, such that Vi = Vj . Let γ be
1 An inducing chain is traditionally called an ‘inducing path’ because that is how it was
originally defined in [Verma, 1992]. However, we use terminology consistent with the rest of
this text.
564
CHAPTER 10. CONSTRAINT-BASED LEARNING
[V1 , . . . Vi ] ⊕ [Vj , . . . Vk ]. Clearly, any cyclic subchain of γ is a cyclic subchain
of ρ, and ρ contains at least one cyclic subchain which γ does not. Therefore γ
contains less cyclic subchains than ρ. It is left as an exercise to show γ is an
inducing chain over V in G between X and Z, and that the edges touching X
and Z on γ have the same direction as the ones touching X and Z on ρ. This
contradiction proves the lemma.
Example 10.16 illustrates the importance of inducing chains. Namely, in
that example, we noticed that the inducing chain between X and Z made it
impossible for any subset of V to d-separate X and Z. This is a general property
of inducing chains. We will give a theorem to that effect shortly. First we need
some lemmas.
Lemma 10.5 Suppose G = (V ∪ H, E) is a hidden node DAG, X, Z ∈ V, and
there is an inducing chain ρ over V in G between X and Z which is out of X
and into Z. Then given any set S ⊆ V − {X, Z}, there exists a chain between
X and Z which is out of X and into Z and which is not blocked by S.
Proof. We have 2 cases:
1. If every collider on ρ either is in S or has a descendent in S, then ρ itself
satisfies the conditions of the lemma.
2. At least one head-to-head on ρ neither is in S nor has a descendent in S.
Let Y be the closest such node to X, and let W be the collider on ρ closest
to X.
If Y = W , Y cannot be an ancestor of X because we would have a directed
cycle. So there is a path λ from Y to Z which does not contain X. Let δ
be ρ(X, Y )⊕ λ. Since δ contains no head-to-head meetings and no nodes
in S, and δ is out of X and into Z, δ satisfies the conditions of the lemma.
If Y 6= W , W is in S or has a descendent in S. Furthermore, since W
is the collider on ρ closest to X, and ρ is out of X, W is a descendent
of X. Therefore, every collider on ρ, which is an ancestor of X, has a
descendent in S. So Y cannot be an ancestor of X, which means there is a
path λ from Y to Z which does not contain X. Since Y has no descendents
in S, λ contains no nodes in S. All colliders on ρ(X, Y ) either are in S or
have a descendent in S. Let δ be ρ(X, Y )⊕ λ. All colliders on δ either are
in S or have a descendent in S, and δ contains no nodes in S which are
not head-to-head. Furthermore, δ is out of X and into Z. So δ satisfies
the conditions of the lemma.
Lemma 10.6 Suppose G = (V ∪ H, E) is a hidden node DAG, X, Z ∈ V, and
there is an inducing chain ρ over V in G between X and Z which is into X and
into Z. Then given any set S ⊆ V − {X, Z}, there exists a chain between X and
Z which is into X and into Z and which is not blocked by S.
Proof. The proof is similar to that of the preceding lemma and is left as an
exercise.
10.2. ASSUMING ONLY EMBEDDED FAITHFULNESS
565
Before proceeding, we introduce more notation. Given a DAG G = (W, E),
X, Z ∈ W, we denote the union of the ancestors of X and the ancestors of Y by
ANCXY .
Lemma 10.7 Suppose G = (V ∪ H, E) is a hidden node DAG, and X, Z ∈ V.
If ρ is a chain between X and Z which is not blocked by ANCXY ∩ V, ρ is an
inducing chain over V in G between X and Z.
Proof. Since ρ is not blocked by ANCXY ∩ V, every collider on ρ is either in
ANCXY ∩ V or is an ancestor of a node in ANCXY ∩ V, which means every
collider on ρ is in ANCXY .
Next we show every node in V, which is an internal node on ρ, is a collider
on ρ. To that end, since every collider on ρ is in ANCXY , every internal node on
ρ is also in ANCXY . This means the set of nodes in V, which are internal nodes
on ρ, is a subset of ANCXY ∩V. However, since ρ is not blocked by ANCXY ∩V,
every member of ANCXY ∩ V on ρ is a collider on ρ. This establishes our
result.
Lemma 10.8 Suppose G = (V ∪ H, E) is a hidden node DAG, and X, Z ∈ V.
Suppose there is an inducing chain ρ over V in G between X and Z which is
out of X. Then there is a path from X to Z. Furthermore, the inducing chain
ρ is into Z.
Proof. If there are no colliders on ρ, then ρ itself is a path from X to Z. If
there are colliders on ρ, let Y be the one closest to X. If Y were the ancestor
of X, we’d have a directed cycle. So Y must be an ancestor of Z. If we let γ be
the path from Y to Z, ρ(X, Y ) ⊕ γ is then a path from X to Z.
As to the second part of the lemma, it is left as an exercise to show we would
have a directed cycle if the inducing chain were also out of Z.
Corollary 10.1 Suppose G = (V ∪ H, E) is a hidden node DAG, and X, Z ∈ V.
Then there is no inducing chain ρ over V in G between X and Z which is both
out of X and out of Z.
Proof. The proof follows immediately from the preceding corollary.
Lemma 10.9 Suppose G = (V ∪ H, E) is a hidden node DAG. Suppose there is
an inducing chain ρ over V in G between X and Z which is out of X. Then
every inducing chain over V in G between X and Z is into Z.
Proof. Owing to Lemma 10.8, there is a path from X to Z. If there were an
inducing chain over V in G between X and Z which were out of Z, that lemma
would also imply there is a path from Z to X. The concatenation of these paths
would be a directed cycle. This contradiction proves the lemma.
Next we give the theorem to which we alluded earlier.
Theorem 10.4 Suppose G = (V ∪ H, E) is a hidden node DAG. For X, Z ∈ V,
X 6= Z, there is an inducing chain over V in G between X and Z if and only if
no subset of V d-separates X and Z.
566
CHAPTER 10. CONSTRAINT-BASED LEARNING
Proof. Suppose there is an inducing chain over V in G between X and Z.
Owing to Lemmas 10.9, 10.5, and 10.6, given any set S ⊆ V − {X, Z}, there
exists a chain between X and Z which is not blocked by S. This means X and
Z are not d-separated by and subset of V.
In the other direction, suppose X and Z are not d-separated by any subset
of V. Then X and Z are not d-separated by ANCXY ∩ V, which means there is
some chain ρ between X and Z which is not blocked by ANCXY ∩ V. Lemma
10.7 therefore implies ρ is an inducing chain over V in G between X and Z.
We close with some further lemmas which will be needed in the next subsection.
Lemma 10.10 Suppose G = (V ∪ H, E) is a hidden node DAG, and there is an
inducing chain ρ over V in G between X and Z which is into Z. Then given
any set S ⊆ V − {X, Z}, there exists a chain between X and Z which is into Z
and which is not blocked by S.
Proof. The proof follows from Lemmas 10.5 and 10.6.
Lemma 10.11 Suppose G = (V ∪ H, E) is a hidden node DAG, X, Y, Z ∈ V,
and we have the following:
1. There is an inducing chain ρ over V in G between X and Z which is into
Z.
2. There is an inducing chain γ over V in G between Z and Y which is into
Z.
Then X and Y are not d-separated by any subset of V containing Z.
Proof. Suppose X and Y are d-separated by S ⊆ V − {X, Y } and Z ∈ S.
Lemma 10.10 says there is a chain δ between X and Z which is into Z and
which is not blocked by S − {Z}, and there is a chain λ between Z and Y which
is into Z and which is not blocked by S − {Z}. Let σ be δ ⊕ λ. The chain σ is
not blocked by S at Z because Z is a collider on σ and Z ∈ S. If σ were blocked
by S somewhere between X and Z, then δ would be blocked by S − {Z}, and if
σ were blocked by S somewhere between Z and Y , then λ would be blocked by
S − {Z}. So σ cannot be blocked anywhere by S, which means X and Y are not
d-separated by S. This contradiction proves the lemma.
Before proceeding, we develop more notation. Given a hidden node DAG G
= (V ∪ H, E), and X, Z ∈ V,
X ← · · · H · · · → Z represents a chain that
1. has no head-to-head meetings;
2. has a tail-to-tail meeting at some node H ∈ H;
X · · · → Z represents a path from X to Z.
10.2. ASSUMING ONLY EMBEDDED FAITHFULNESS
X
567
Z
(a)
X
H
Z
(b)
X
Z
(c)
Figure 10.17: If there is an inducing chain ρ over V in G between X and Z, we
must have one of these situations.
Lemma 10.12 Suppose G = (V ∪ H, E) is a hidden node DAG, X, Z ∈ V, and
there is an inducing chain over V in G between X and Z which is into Z. Then
at least one of the following holds:
a) There is a path X · · · → Z.
b) There is a simple chain X ← · · · H · · · → Z.
The possibilities are depicted in Figure 10.17 (a) and (b).
Proof. Owing to Lemma 10.4 there is a simple inducing chain ρ over V in G
between X and Z which is into Z. We have 3 cases:
1. There are no head-to-head meetings on ρ. In this case ρ itself satisfies
property (a) or (b).
2. There are head-to-head meetings on ρ, and no colliders on ρ are ancestors
of Z. Let Y be the collider closest to Z on ρ. There is a path γ from Y
to X which does not go through Z. Note that this path must be simple
because it is in a DAG. Let W be the node on γ which is closest to Z on
ρ. Then γ(X, W ) ⊕ ρ(W, Z) satisfies property (b).
3. There are head-to-head meetings on ρ, and at least one collider is an ancestor of Z. Let Y be the closest such node to X. There is a path γ from Y
to Z. If γ goes through X, then there is a path from X to Z and property
(a) is satisfied. So suppose γ does not go through X. Let W be the node
on γ which is closet to X on ρ. If there are no head-head nodes between
X and W on ρ, the chain ρ(X, W ) ⊕ γ(W, Z) satisfies either property (a)
or (b). If there are head-head nodes between X and W on ρ, let T be the
one closest to W on ρ. There is a path δ from T to X which does not
go through Z. Let U be the node on δ which is closest to W on ρ. Then
chain δ(X, U ) ⊕ ρ(U, W ) ⊕ γ(W, Z) satisfies property (b).
568
CHAPTER 10. CONSTRAINT-BASED LEARNING
Lemma 10.13 Suppose G = (V ∪ H, E) is a hidden node DAG, X, Z ∈ V, and
there is an inducing chain ρ over V in G between X and Z. Then at least one
of the following holds:
a) There is a path X · · · → Z.
b) There is a simple chain X ← · · · H · · · → Z.
c) There is a path Z · · · → X.
The possibilities are depicted in Figure 10.17.
Furthermore, (a) and (c) cannot both hold.
Proof. The first part of the lemma follows from Corollary 10.1 and Lemma
10.12. As to (a) and (c) both holding, we would have a directed cycle if this
were the case.
Lemma 10.14 Suppose G = (V ∪ H, E) is a hidden node DAG, X, Z ∈ V, there
is an inducing chain over V in G between X and Z which is into Z, and there
is an inducing chain over V in G between X and Z which is into X. Then the
following holds:
There is a simple chain X ← · · · H · · · → Z.
Proof. Suppose there is no such simple chain. Then, due to Lemma 10.12,
there is a path λ from X to Z, and there is a path γ from Z to X. Since λ ⊕ γ
is a directed cycle, we have a contradiction.
10.2.2
A Basic Algorithm
Next we develop an algorithm which determines common features of every DAG
in which a set of d-separations is embedded faithfully. First we discuss the links
and notation used in the algorithm:
• There is a new kind of link X ½ Z.
• It is possible for the algorithm to produce the links X → Z and X ← Z.
When this is the case, we write X ↔ Z.
• Some meetings are marked as follows: X−Z−Y . They are called marked
meetings.
• The symbol ∗ is used to denote any of the possible edge endings that may
be touching the node. For example,
X∗ → Y means the link could be X → Y , X ½ Y , or X ↔ Y .
• By a directed path, we mean a path in which all the links are of the
form X ½ Z.
10.2. ASSUMING ONLY EMBEDDED FAITHFULNESS
569
We prove theorems concerning the meaning of the links after presenting the
algorithm. The algorithm produces a graph, which we call a hidden node
DAG pattern. Later we will formally define this term.
Algorithm 10.5 Learn Embedded Faithfulness
Problem: Given a set IND of d-separations, determine common features of
every DAG in which IND is embedded faithfully.
Inputs: a set V of nodes and a set IND of d-separations amongst the nodes.
Outputs: If IND admits an embedded faithful DAG representation, a hidden
node DAG pattern gp showing common features of every DAG in which
IND is embedded faithfully.
void learn_embedded (set-of-nodes V,
set-of-d-separations IND,
graph& gp)
{
for (each pair of nodes X, Y ∈ V) {
search for a subset SXY ⊆ V such that
I({X}, {Y }|SXY ) ∈ IND;
if (no such set can be found) {
create the link X − Y in gp;
}
for (each uncoupled meeting X − Z − Y )
if (Z ∈
/ SXY )
orient X − Z − Y as X → Z ← Y ;
else
mark X − Z − Y as X−Z−Y ;
while (more edges can be oriented) {
for (each marked meeting X∗ → Z − ∗Y );
orient Z − ∗Y as Z ½ Y ;
for (each link X ∗ −Y )
if (there is a directed path from X to Y )
orient X ∗ −Y as X∗ → Y ;
for (each head-to-head meeting X∗ → W ← ∗Y )
if (Z ∗ −W && Z ∈ SXY )
orient Z ∗ −W as Z∗ → W ;
}
// Step 1
// Step 2
// Step 3
// Step 4
// Step 5
}
As shown by the following lemmas and theorems, the graph, produced by
the preceding algorithm, tells us common features of every DAG in which IND
is embedded faithfully.
570
CHAPTER 10. CONSTRAINT-BASED LEARNING
Prp.
1
2
3
Link
X and Y
not linked
X −Y
X→Y
4
X½Y
5
X↔Y
6
X∗ − ∗Z ∗ −∗Y
Meaning (inducing chains are over V in G = (V ∪ H, E))
No inducing chain between X and Y .
Inducing chain between X and Y .
Inducing chain between X and Y into Y ,
and no path Y · · · → X.
Inducing chain between X and Y out of X/into Y ,
and no inducing chain between X and Y into X.
Inducing chain between X and Y into X/into Y ,
and no path Y · · · → X,
and no path X · · · → Y .
Inducing chain between Z and X out of Z/into X,
and no inducing chain between Z and X into Z;
or
inducing chain between Z and Y out of Z/into Y ,
and no inducing chain between Z and Y into Z.
Table 10.1: The meaning of the links created by Algorithm 10.5 for any DAG
in which a set of d-separations is embedded faithfully. The ‘or’ is inclusive.
Lemma 10.15 Suppose IND is a set of d-separations amongst a set of nodes
V, and IND is embedded faithfully in DAG G = (V ∪ H, E). If IND is the input
to Algorithm 10.5, that algorithm creates a link between X and Z if and only if
no subset of V d-separates X and Z in G.
Proof. Since IND is embedded faithfully in DAG G, all and only the d-separations
in G restricted to elements of V are in IND. Algorithm 10.5 creates a link between X and Z if and only if there is no d-separation in IND for X and Z,
which means no subset of V d-separates X and Z in G.
Lemma 10.16 Suppose IND is a set of d-separations amongst a set of nodes
V, and IND is embedded faithfully in DAG G = (V ∪ H, E). If IND is the input
to Algorithm 10.5, that algorithm creates a link between X and Z if and only if
there is an inducing chain over V in G between X and Z.
Proof. The proof follows from Lemma 10.15 and Theorem 10.4.
Theorem 10.5 Suppose IND is a set of d-separations amongst a set of nodes V,
and IND admits an embedded faithful DAG representation. If IND is the input
to Algorithm 10.5, the links created by the algorithm have the meaning shown
in Table 10.1 for any DAG G = (V ∪ H, E) in which IND is embedded faithfully.
Proof. Let G be a DAG in which IND is embedded faithfully. We prove each
of the properties in turn:
• Properties 1 and 2 were proven in Lemma 10.16.
• Properties 3 and 4 will be established using induction on the links created
by the algorithm.
10.2. ASSUMING ONLY EMBEDDED FAITHFULNESS
571
Induction base : The initial arrowheads created are due to uncoupled meet/ SXY . If
ings like X → Z ← Y , where I({X}, {Y }|SXY ) ∈ IND, and Z ∈
no such arrowheads are created, then the algorithm will create no arrowheads and we are done. So assume k such arrowheads are created where
k ≥ 2. Lemma 10.16 says there is an inducing chain ρ over V in G
between X and Z, and there is an inducing chain δ over V in G between
Z and Y . Suppose there is a path Z · · · → X in G. Then all ancestors of
Z are also ancestors of X, which means ρ⊕ δ is an inducing chain over
V −{Z} in G between X and Y . Theorems 2.5 and 10.4 therefore imply no subset of V −{Z} can render X and Y conditionally independent.
This contradiction establishes there can be no path Z · · · → X in G, and
therefore Lemma 10.8 implies there can be no inducing chain over V in G
between X and Z which is out of Z. So ρ is into Z. Similarly, there can
be no path Z · · · → Y and δ is into Z.
Induction hypothesis : Suppose for the first m links created, where m ≥ k,
the links X → Y and X ½ Y have the meaning in Table 10.1.
Induction step : Consider the (m + 1)st link created. We consider each of
the possible steps that could create it in turn:
Suppose Step 3 creates the link Z ½ Y . The link is created because there
is a marked meeting X∗ → Z − ∗Y , which means it is an uncoupled
meeting X∗ → Z − ∗Y . Due to the induction hypothesis, there is an
inducing chain ρ over V in G between X and Z which is into Z. Since
Algorithm 10.5 did not create a link between X and Y , there is some
subset SXY ⊆ V that renders them conditionally independent. If Z
were not in SXY , Algorithm 10.5 would have oriented X → Z ← Y .
So Z ∈ SXY . Therefore, due to Theorem 2.5, SXY is a set containing
Z which d-separates X and Y in G. Lemma 10.11 therefore says there
cannot be an inducing chain over V in G between Z and Y , which is
into Z, which proves the second part of the property. Lemma 10.16
says there is some inducing chain γ over V in G between Z and Y .
So γ must be out of Z. The first part of the property now follows
from Lemma 10.8.
Suppose Step 4 creates the link X∗ → Y . Since there is a directed path
from X to Y in the pattern created by Algorithm 10.5, due to the induction hypothesis and Lemma 10.8, there is a path X · · · → Y in G.
This implies, again due to Lemma 10.8, there cannot be an inducing
chain over V in G between X and Y which is out of Y . Lemma 10.16
says there is some inducing chain γ over V in G between X and Y .
So γ must be into Y . Finally, if there were a path Y · · · → X in G,
we’d have a directed cycle in G.
Suppose Step 5 creates the link Z∗ → W . We know that I({X}, {Y }|SXY )
where Z ∈ SXY . This means, due to Theorem 2.5, every chain
in G between X and Y must be blocked by SXY . Since we have
X∗ → W ← ∗Y , due to the induction hypothesis and Lemma 10.10,
572
CHAPTER 10. CONSTRAINT-BASED LEARNING
there is a chain ψ in G between X and W which is into W , a chain δ
in G between Y and W which is into W , and neither chain is blocked
by SXY . Consider the chain ψ ⊕ δ between X and Y . Clearly it cannot be blocked by SXY at an interior node on ψ or on δ. Therefore,
it must be blocked at W . If there were a path W · · · → Z in G, then
ψ ⊕ δ would not be blocked by SXY at W because Z ∈ SXY . We
conclude there can be no path W · · · → Z in G, and due to Lemma
10.8, no inducing chain over V in G between W and Z, which is out
of W . However, Lemma 10.16 says there is some inducing chain γ
over V in G between Z and W . So γ is into W .
• Next consider Property 5. Such a link X ↔ Y means the algorithm created
the links X → Y and X ← Y . First we show there is an inducing chain
over V in G between X and Y which is into X and into Y . Owing to
Property 3, we know there is an inducing chain ρ over V in G between X
and Y which is into Y , and an inducing chain γ over V in G between X
and Y which is into X. Suppose ρ is out of X and γ is out of Y . Then
Lemma 10.8 implies there is a path from X to Y in G and a path from Y
to X in G. But in this case we would have a directed cycle. Therefore,
either ρ or γ must be both into X and into Y . As to the second part of
Property 5, the fact that there can be no such paths follows directly from
Property 3.
• Finally, consider Property 6. Such a marked meeting is created only if
we have I({X}, {Y }|SXY ) where Z ∈ SXY . Suppose there is an inducing
chain over V in G between X and Z which is into Z, and an inducing
chain over V in G between Z and Y which is into Z. Then due to Lemma
10.11, X and Y are not d-separated by any subset of V containing Z. This
contradictions shows there is either no inducing chain over V in G between
X and Z which is into Z, or no inducing chain over V in G between Z
and Y which is into Z. Without loss of generality, suppose the former
is true. Lemma 10.16 says there is some inducing chain ρ over V in G
between X and Z. So ρ must be out of Z, and due to Lemma 10.8 ρ must
be into X, which completes the proof.
Next we give a theorem concerning higher level meaning of the links created
by Algorithm 10.5.
Theorem 10.6 Suppose IND is a set of d-separations amongst a set of nodes V,
and IND admits an embedded faithful DAG representation. If IND is the input to
Algorithm 10.5, the links created by the algorithm have the higher level meaning
shown in Table 10.2 for any DAG G = (V ∪ H, E) in which IND is embedded
faithfully.
Proof. We prove each property in turn:
1. If a chain, which contains only hidden nodes on its interior, did not have
a head-to-head meeting on it, it would be an inducing chain over V in G
10.2. ASSUMING ONLY EMBEDDED FAITHFULNESS
Prp.
1
2
Link
X and Y
not linked
X −Y
3
X→Y
4
X½Y
5
X↔Y
6
X∗ − ∗Z ∗ −∗Y
573
Meaning
Any chain between X and Y containing only hidden
nodes has a head-to-head meeting on it.
Path X · · · → Y (exclusive) or path Y · · · → X;
or simple chain X ← · · · H · · · → Y .
Path X · · · → Y or simple chain X ← · · · H · · · → Y ;
and no path Y · · · → X.
Path X · · · → Y and no chain
X ← · · · H · · · → Y containing only hidden nodes.
Simple chain X ← · · · H · · · → Y
and no path Y · · · → X
and no path X · · · → Y .
Path Z · · · → X and no chain
X ← · · · H · · · → Z containing only hidden nodes;
or
path Z · · · → Y and no chain
Z ← · · · H · · · → Y containing only hidden nodes.
Table 10.2: The high level meaning of the links created by Algorithm 10.5 for
any DAG in which a set of d-separations is embedded faithfully. When we say a
chain ‘contains only hidden nodes,’ we mean on its interior. ‘Or’s are inclusive
unless denoted as exclusive.
between X and Y . However, Property 1 in Theorem 10.5 says there can
be no inducing chain over V in G between X and Y .
2. Property 2 follows from Property 2 in Theorem 10.5 and Lemma 10.13.
3. Property 3 follows from Property 3 in Theorem 10.5 and Lemma 10.12.
4. The first part of Property 4 follows from the first part of Property 4 in
Theorem 10.5 and Lemma 10.8. As to the second part, the second part of
Property 4 in Theorem 10.5 says there can be no inducing chain over V
between X and Y which is into X. However, a chain X ← · · · H · · · → Y ,
containing only hidden nodes on its interior, is such an inducing chain.
5. Property 5 follows from Property 5 in Theorem 10.5 and Lemma 10.14.
6. The proof of Property 6 is just like that of Property 4.
Next we give some examples.
Example 10.17 Suppose V = {N, F, C, T } and IND consists of the d-separations
in the DAG in Figure 10.18 (a) restricted to these nodes. Given this input, Algorithm 10.5 will produce the hidden node DAG pattern in Figure 10.18 (b).
From Table 10.2, we see IND could not be embedded faithfully in the DAGs in
10.18 (e) and (f). It is left as an exercise to show IND is embedded faithfully in
574
CHAPTER 10. CONSTRAINT-BASED LEARNING
N
T
H
F
N
C
T
F
(a)
N
H2
(b)
T
H1
F
N
H2
C
F
C
(d)
T
H
F
T
H1
(c)
N
C
C
(e)
N
T
H
F
C
(f)
Figure 10.18: Given V = {N, F, C, T } and IND consists of the d-separations
in the DAG in (a) restricted to these nodes, Algorithm 10.5 will produce the
hidden node DAG pattern in (b). IND is embedded faithfully in the DAGs in
(a), (c) and (d), but it is not embedded faithfully in the DAGs in (e) and (f).
the DAGs in Figures (a), (c), and (d). Of course, it is embedded faithfully in
many more DAGs (indeed an infinite number).
Example 10.18 Suppose V = {N, F, C, T, A, D} and IND consists of the dsepara-tions in the DAG in Figure 10.19 (a) restricted to these nodes. Given
this input, Algorithm 10.5 will produce the hidden node DAG pattern in Figure
10.19 (b). In this and all future figures, we do not show marked meetings what
are apparent from the arrowheads. For example, the meeting N → F ½ D
is marked as N → F ½D, but we do not show this in the figure because it is
apparent from the link F ½ D.
Owing to Property 6 in Table 10.2, in any DAG in which IND is embedded
faithfully, either
there is a path A · · · → N and
there is no simple chain A ← · · · H · · · → N containing only hidden nodes on
its interior;
10.2. ASSUMING ONLY EMBEDDED FAITHFULNESS
A
N
A
H
F
575
T
C
D
N
T
F
C
D
(a)
(b)
Figure 10.19: Given V = {N, F, C, T, A, D} and IND consists of the dseparations in the DAG in (a) restricted to these nodes, Algorithm 10.5 will
produce the hidden node DAG pattern in (b).
or
there is a path A · · · → T and
there is no simple chain A ← · · · H · · · → T containing only hidden nodes on
its interior.
Furthermore, if, for example, we have the first (i.e. a path A · · · → N , etc.),
then, again owing to Property 6 in Table 10.2, we know
There is a path N · · · → F and
there is no simple chain N ← · · · H · · · → F containing only hidden nodes on
its interior.
Example 10.19 Suppose V = {X, Y, Z, W } and IND consists of the d-separations in the DAG in Figure 10.20 (a) restricted to these nodes. Given this input,
Algorithm 10.5 will produce the hidden node DAG pattern in Figure 10.20 (b).
It is not hard to see that IND is embedded faithfully in the DAGs in Figures
10.20 (c) and (d). This example shows that, when Algorithm 10.5 creates an
edge X → Z, Property 3 in Table 10.2 does not entail there must be a path
X · · · → Z or simple chain X ← · · · H · · · → Z containing only hidden nodes
on its interior (or no nodes). In Figure 10.20 (c) we have only the path X →
W → Z satisfying that property, and in Figure 10.20 (d) we have only the simple
chain X ← H1 → W → Z satisfying that property.
576
CHAPTER 10. CONSTRAINT-BASED LEARNING
X
W
Z
X
W
Z
Y
Y
(a)
(b)
H
X
W
H2
H1
Z
X
W
Z
Y
Y
(c)
(d)
Figure 10.20: Given V = {X, Y, Z, W } and IND consists of the d-separations
in the DAG in (a) restricted to these nodes, Algorithm 10.5 will produce the
hidden node DAG pattern in (b). IN DG is embedded faithfully in the DAGs
in (c) and (d).
U
U
X
Y
X
Z
Y
Z
V
V
(b)
(a)
U
H
X
Y
Z
V
(c)
Figure 10.21: Given V = {X, Y, Z, U, V } and IND consists of the d-separations
in the DAG in (a) restricted to these nodes, Algorithm 10.5 will produce the
hidden node DAG pattern in (b). IND is embedded faithfully in the DAG in
(c).
10.2. ASSUMING ONLY EMBEDDED FAITHFULNESS
W
H1
X
H2
Y
577
T
Z
(a)
W
T
X
Y
Z
(b)
Figure 10.22: Given V = {X, Y, Z, W, T } and INDG consists of the d-separations
in the DAG in (a) restricted to these nodes, Algorithm 10.5 will produce the
hidden node DAG pattern in (b).
Example 10.20 Suppose V = {X, Y, Z, U, V } and IND consists of the d-separations in the DAG in Figure 10.21 (a) restricted to these nodes. Given this input,
Algorithm 10.5 will produce the hidden node DAG pattern in Figure 10.21 (b). It
is not hard to see that IND is embedded faithfully in the DAG in Figure 10.21 (c).
This example shows that, when Algorithm 10.5 creates an edge X ½ Z, Property
4 in Table 10.2 does not entail there must be a path X · · · → Z containing only
hidden nodes on its interior (or no nodes). In Figure 10.21 (c) we have only
the path X → Y → Z satisfying that property.
Example 10.21 Suppose V = {X, Y, Z, W, T } and IND consists of the d-separations in the DAG in Figure 10.22 (a) restricted to these nodes. Given this input,
Algorithm 10.5 will produce the hidden node DAG pattern in Figure 10.22 (b).
Clearly IND is embedded faithfully in the DAG in Figure 10.22 (a). This example
shows that, when Algorithm 10.5 creates an edge X ↔ Z, Property 5 in Table
10.2 does not entail there must be a chain X ← · · · H · · · → Z containing
only hidden nodes on its interior. In Figure 10.22 (a) we have only the chain
X ← H1 → Y → Z satisfying that property.
Example 10.22 Suppose V = {X, Y, Z, U, V, W } and IND consists of the dsepara-tions in the DAG in Figure 10.23 (a) restricted to these nodes. Given
this input, Algorithm 10.5 will produce the hidden node DAG pattern in Figure
10.23 (b). It is not hard to see that IND is embedded faithfully in the DAG in
Figure 10.23 (c). This example shows that, when Algorithm 10.5 creates an edge
578
CHAPTER 10. CONSTRAINT-BASED LEARNING
W
U
Y
X
W
U
Z
V
Y
X
Z
V
(a)
(b)
H
U
W
Y
X
Z
V
(c)
Figure 10.23: Given V = {X, Y, Z, U, V, W } and IND consists of the dseparations in the DAG in (a) restricted to these nodes, Algorithm 10.5 will
produce the hidden node DAG pattern in (b). IND is embedded faithfully in
the DAG in (c).
X ½ Z, there can still be a chain X ← · · · H · · · → Z which does not contain
only hidden nodes on its interior.
Example 10.23 Suppose V = {X, Y, Z, V, W } and IND consists of the d-separations in the DAG in Figure 10.24 (a) restricted to these nodes. Given this input, Algorithm 10.5 will produce the hidden node DAG pattern in Figure 10.24
(b).This example shows that it is not correct to direct arrows to avoid directed
cycles when arrows in the path are of the form →. When Step 4 is encountered,
Step 5 has not yet directed Y → X as Y ←→ X. So, if we directed Z − X
as Z → X due to the path Z → Y → X, it would mean IND is not embedded
faithfully in any DAG with a path from X to Z. Clearly, the DAG in Figure
10.24 (a) is such a DAG.
An Improved Algorithm
Next we develop an algorithm which learns more about every DAG in which
IND is embedded faithfully than Algorithm 10.5. First we develop the algorithm;
then we improve its efficiency.
10.2. ASSUMING ONLY EMBEDDED FAITHFULNESS
V
Z
H1
Y
579
W
H2
X
(a)
Z
V
W
Y
X
(b)
Figure 10.24: Given V = {X, Y, Z, V, W } and IND consists of the d-separations
in the DAG in (a) restricted to these nodes, Algorithm 10.5 will produce the
hidden node DAG pattern in (b).
The Algorithm
We start by formally defining ‘hidden node DAG pattern’.
Definition 10.2 A hidden node DAG pattern is a graph with links and
marked meetings like those shown in Table 10.1.
Definition 10.3 Suppose IND is a set of d-separations amongst a set V of nodes
which admits an embedded faithful DAG representation. We say IND is embedded in hidden node DAG pattern gp if the links in gp have the meaning shown
in Table 10.1 (and therefore the high level meaning shown in Table 10.2) for
any DAG G in which IND is embedded faithfully.
It is clear that Algorithm 10.5 creates a hidden node DAG pattern in which
IND is embedded.
Consider two hidden node DAG patterns gp1 and gp2 in which IND is embedded. Suppose there is no link between X and Y in gp1 . Then owing to
Property 1 in Table 10.1, there is no inducing chain over V between X and Y
in any DAG in which IND is embedded faithfully. If there were a link between
X and Y in gp2 , then owing to Properties 2-6 in Table 10.1, there would be an
inducing chain over V between X and Y in any DAG in which IND is embedded
faithfully. So there can be no link between X and Y in gp2 . We conclude all
580
CHAPTER 10. CONSTRAINT-BASED LEARNING
N
T
H
F
N
C
T
F
(a)
(b)
N
T
F
N
C
T
F
(c)
C
(d)
N
T
F
C
C
N
H2
(e)
T
H1
F
C
(f)
Figure 10.25: Given V = {N, F, C, T } and IND consists of the d-separations
in the DAG in (a) restricted to these nodes, IND is embedded in the hidden
node DAG patterns in (b), (c), and (d), and IND is maximally embedded in
the hidden node DAG pattern in (b). IND is not embedded in the hidden node
DAG pattern in (e).
hidden node DAG patterns in which IND is embedded have the same links. The
difference is in the arrowheads and marked meetings.
Definition 10.4 Suppose IND is a set of d-separations amongst a set V of nodes
which admits an embedded faithful DAG representation. We say IND is maximally embedded in hidden node DAG pattern gp if IND is embedded in gp,
and there is no hidden node DAG pattern, in which IND is embedded, which has
either
1. more links marked with arrowheads (either at the beginning or end); or
2. more marked meetings X∗ − ∗Z ∗ −∗Y.
Example 10.24 Suppose V = {N, F, C, T } and IND consists of the d-separations
in the DAG in Figure 10.25 (a) restricted to these nodes. Then IND is embedded in the hidden node DAG patterns in Figure 10.25 (b), (c), and (d). Clearly,
IND is not maximally embedded in the patterns Figure 10.25 (c) or (d). Is IND
10.2. ASSUMING ONLY EMBEDDED FAITHFULNESS
581
maximally embedded in the pattern in Figure 10.25 (b)? If IND were not, there
would have to be a hidden node DAG pattern, in which IND is embedded, with
more arrows than the one in Figure 10.25 (b). That pattern would have to have
N ½ F or T ½ C. A pattern with N ½ F is shown in Figure 10.25 (e).
However, as shown in Example 10.17, IND is embedded faithfully in the DAG in
Figure 10.25 (f). This DAG does not have a path N · · · → F , and it has a chain
N ← · · · H · · · → F containing only hidden nodes. Table 10.2 says neither of
these are allowed if we have N ½ F . Similarly, we cannot have T ½ C. So
IND is maximally embedded in the hidden node DAG pattern in Figure 10.25
(b).
As shown in Example 10.17, Algorithm 10.5 produces the hidden variable
DAG pattern in Figure 10.25 (b) when IND consists of the d-separations in the
DAG in Figure 10.25 (a). So in this case Algorithm 10.5 created a hidden node
DAG pattern in which IND is maximally embedded. Is this always the case?
Consider the following example.
Example 10.25 Suppose V = {X, Y, Z, U, V, W } and IND consists of the dseparations in the DAG in Figure 10.26 (a) restricted to these nodes. Given
this input, Algorithm 10.5 will produce the hidden node DAG pattern gp in
Figure 10.26 (b). Next we show IND is not maximally embedded in that pattern.
To that end, let G be a DAG in which IND is embedded faithfully. Suppose there
is a path ρ = Z ← · · · Y in G. SinceIND is embedded in gp, owing to Table
10.2 the link X ½ W in gp means we must have a path γ = X · · · → W in G,
and the link W ½ Y in gp means we must have a path δ = W · · · → Y in G .
Consider the chain γ ⊕ δ ⊕ ρ between X and Z. This chain is not blocked by ∅.
Yet, since IND is embedded faithfully in the DAG in Figure 10.26 (a), clearly we
have I({X}, {Z}). This contradiction shows there can be no such path ρ. Due to
Table 10.2, we can therefore orient the link Z − Y in gp as Z → Y . It is left as
an exercise to show that we can also mark W ← Z → Y in gp as W ← Z →Y .
So IND is embedded in the hidden node DAG pattern in Figure 10.26 (c). It is
also left as an exercise to show IND is maximally embedded in this pattern.
Based on considerations like those in Example 10.25, Spirtes et al [1993,2000]
developed an algorithm which creates more arrowheads and marked meetings
than Algorithm 10.5. The algorithm requires the concept of a definite discriminating chain, which is defined shortly. First we need some preliminary
definitions. Let gp be a hidden node DAG pattern, ρ be a chain in gp, Z be an
interior node on ρ, and X and Y the nodes adjacent to Z on ρ. Then Z is called
a collider on ρ if we have X∗ → Z ← ∗Y on ρ, and Z is called a definite
non-collider on ρ if we have X∗ − ∗Z ∗ −∗Y on ρ.
Definition 10.5 Let gp be a hidden node DAG pattern and ρ be chain between
X and Y . Then ρ is a definite discriminating chain for Z in gp if the
following hold:
1. X and Y are not adjacent in gp.
582
CHAPTER 10. CONSTRAINT-BASED LEARNING
H
U
X
W
Z
Y
Z
Y
Z
Y
V
(a)
U
X
W
V
(b)
U
X
W
V
(c)
Figure 10.26: Given V = {X, Y, Z, U, V, W } and IND consists of the dseparations in the DAG in (a) restricted to these nodes, Algorithm 10.5, will
produce the hidden node DAG pattern in (b). IND is maximally embedded in
the hidden node DAG pattern in (c).
2. Z is an interior node on ρ.
3. Every interior node on ρ, except for Z, is a collider or definite noncollider.
4. If U and V are adjacent on ρ and V is between U and Z on ρ, then
U ∗ →V on ρ.
5. If V is between X and Z, then
if V is a collider on ρ, then
V ½ Y on ρ
else
V ← ∗Y on ρ.
6. If V is between Y and Z, then
10.2. ASSUMING ONLY EMBEDDED FAITHFULNESS
583
if V is a collider on ρ, then
V ½ X on ρ
else
V ← ∗X on ρ.
Example 10.26 Consider the hidden node DAG pattern gp in Figure 10.26
(b). The chain X ½ W ½ Y − Z is a definite discriminating chain for Y in
gp. The chain X ½ W ← Z − Y is a definite discriminating chain for Z in gp.
In Example 10.25 it was the definite discriminating chain X ½ W ½ Y − Z
for Y in gp that enabled us to conclude we could orient the link Z − Y in gp as
Z → Y . Note that we have I({X}, {Z}) and Y ∈
/ ∅. Furthermore, it was the
definite discriminating chain X ½ W ← Z − Y for Z in gp that enabled us to
mark W ← Z → Y in gp as W ← Z →Y . Note that we have I({X}, {Y }|{W, Z})
and Z ∈ {W, Z}. These situations illustrate general properties of discriminating
chains. Specifically, we have the following theorem:
Theorem 10.7 Suppose IND is a set of d-separations amongst a set V of nodes
which admits an embedded faithful DAG representation. Suppose further gp is a
hidden node DAG pattern in which IND is embedded, and ρ is a chain between
X and Y in gp, which is a definite discriminating chain for Z in gp. Let U
and V be the nodes adjacent to Z on ρ, and let SXY be a subset of V such
that I({X}, {Y }|SXY ). Then IND is embedded in hidden node DAG pattern
gp0 , where gp0 has the same arrowheads and marked meeting as gp and also the
following ones:
1. U ∗ − ∗Z ∗ −∗V
if Z ∈ SXY .
2. U ∗ → Z ← ∗V
if Z ∈
/ SXY .
Proof. The proof can be found in [Spirtes et al, 1993, 2000].
Example 10.27 Suppose again V = {X, Y, Z, U, V, W } and IND consists of
the d-separations in the DAG in Figure 10.26 (a) restricted to these nodes.
Recall that if we use this distribution as the input to Algorithm 10.5, we obtain
the hidden node DAG pattern gp in Figure 10.26 (b), and the chain X ½
W ½ Y − Z is a definite discriminating chain for Y in gp. Clearly, we have
I({X}, {Z}). Owing to Theorem 10.7, IND is embedded in the hidden node
DAG pattern gp0 obtained from gp by marking W ½ Y − Z as W ½ Y ← Z.
Similarly, since X ½ W ← Z → Y is a definite discriminating chain for Z
in gp0 and we have I({X}, {Y }|{W, Z}), we can then mark W ← Z → Y as
W ← Z →Y . The final pattern appears in Figure 10.26 (c).
Next we give the improvement to Algorithm 10.5. The c
Download